prometheus/storage/remote/wal_watcher.go

579 lines
16 KiB
Go
Raw Normal View History

Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
// Copyright 2018 The Prometheus Authors
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
package remote
import (
"fmt"
"os"
"path"
"strconv"
"strings"
"time"
"github.com/go-kit/kit/log"
"github.com/go-kit/kit/log/level"
"github.com/pkg/errors"
"github.com/prometheus/client_golang/prometheus"
"github.com/prometheus/tsdb"
"github.com/prometheus/tsdb/fileutil"
"github.com/prometheus/tsdb/wal"
)
const (
readPeriod = 10 * time.Millisecond
checkpointPeriod = 5 * time.Second
segmentCheckPeriod = 100 * time.Millisecond
)
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
var (
watcherSamplesRecordsRead = prometheus.NewCounterVec(
prometheus.CounterOpts{
Namespace: "prometheus",
Subsystem: "wal_watcher",
Name: "samples_records_read_total",
Help: "Number of samples records read by the WAL watcher from the WAL.",
},
[]string{queue},
)
watcherSeriesRecordsRead = prometheus.NewCounterVec(
prometheus.CounterOpts{
Namespace: "prometheus",
Subsystem: "wal_watcher",
Name: "series_records_read_total",
Help: "Number of series records read by the WAL watcher from the WAL.",
},
[]string{queue},
)
watcherTombstoneRecordsRead = prometheus.NewCounterVec(
prometheus.CounterOpts{
Namespace: "prometheus",
Subsystem: "wal_watcher",
Name: "tombstone_records_read_total",
Help: "Number of tombstone records read by the WAL watcher from the WAL.",
},
[]string{queue},
)
watcherInvalidRecordsRead = prometheus.NewCounterVec(
prometheus.CounterOpts{
Namespace: "prometheus",
Subsystem: "wal_watcher",
Name: "invalid_records_read_total",
Help: "Number of invalid records read by the WAL watcher from the WAL.",
},
[]string{queue},
)
watcherUnknownTypeRecordsRead = prometheus.NewCounterVec(
prometheus.CounterOpts{
Namespace: "prometheus",
Subsystem: "wal_watcher",
Name: "unknown_records_read_total",
Help: "Number of records read by the WAL watcher from the WAL of an unknown record type.",
},
[]string{queue},
)
watcherRecordDecodeFails = prometheus.NewCounterVec(
prometheus.CounterOpts{
Namespace: "prometheus",
Subsystem: "wal_watcher",
Name: "record_decode_failures_total",
Help: "Number of records read by the WAL watcher that resulted in an error when decoding.",
},
[]string{queue},
)
watcherSamplesSentPreTailing = prometheus.NewCounterVec(
prometheus.CounterOpts{
Namespace: "prometheus",
Subsystem: "wal_watcher",
Name: "samples_sent_pre_tailing_total",
Help: "Number of sample records read by the WAL watcher and sent to remote write during replay of existing WAL.",
},
[]string{queue},
)
watcherCurrentSegment = prometheus.NewGaugeVec(
prometheus.GaugeOpts{
Namespace: "prometheus",
Subsystem: "wal_watcher",
Name: "current_segment",
Help: "Current segment the WAL watcher is reading records from.",
},
[]string{queue},
)
)
func init() {
prometheus.MustRegister(watcherSamplesRecordsRead)
prometheus.MustRegister(watcherSeriesRecordsRead)
prometheus.MustRegister(watcherTombstoneRecordsRead)
prometheus.MustRegister(watcherInvalidRecordsRead)
prometheus.MustRegister(watcherUnknownTypeRecordsRead)
prometheus.MustRegister(watcherRecordDecodeFails)
prometheus.MustRegister(watcherSamplesSentPreTailing)
prometheus.MustRegister(watcherCurrentSegment)
}
type writeTo interface {
Append([]tsdb.RefSample) bool
StoreSeries([]tsdb.RefSeries, int)
SeriesReset(int)
}
// WALWatcher watches the TSDB WAL for a given WriteTo.
type WALWatcher struct {
name string
writer writeTo
logger log.Logger
walDir string
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
currentSegment int
lastCheckpoint string
startTime int64
samplesReadMetric prometheus.Counter
seriesReadMetric prometheus.Counter
tombstonesReadMetric prometheus.Counter
invalidReadMetric prometheus.Counter
unknownReadMetric prometheus.Counter
recordDecodeFailsMetric prometheus.Counter
samplesSentPreTailing prometheus.Counter
currentSegmentMetric prometheus.Gauge
quit chan struct{}
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
}
// NewWALWatcher creates a new WAL watcher for a given WriteTo.
func NewWALWatcher(logger log.Logger, name string, writer writeTo, walDir string, startTime int64) *WALWatcher {
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
if logger == nil {
logger = log.NewNopLogger()
}
w := &WALWatcher{
logger: logger,
writer: writer,
walDir: path.Join(walDir, "wal"),
startTime: startTime,
quit: make(chan struct{}),
}
w.samplesReadMetric = watcherSamplesRecordsRead.WithLabelValues(name)
w.seriesReadMetric = watcherSeriesRecordsRead.WithLabelValues(name)
w.tombstonesReadMetric = watcherTombstoneRecordsRead.WithLabelValues(name)
w.unknownReadMetric = watcherUnknownTypeRecordsRead.WithLabelValues(name)
w.invalidReadMetric = watcherInvalidRecordsRead.WithLabelValues(name)
w.recordDecodeFailsMetric = watcherRecordDecodeFails.WithLabelValues(name)
w.samplesSentPreTailing = watcherSamplesSentPreTailing.WithLabelValues(name)
w.currentSegmentMetric = watcherCurrentSegment.WithLabelValues(name)
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
return w
}
func (w *WALWatcher) Start() {
level.Info(w.logger).Log("msg", "starting WAL watcher", "queue", w.name)
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
go w.runWatcher()
}
func (w *WALWatcher) Stop() {
level.Info(w.logger).Log("msg", "stopping WAL watcher", "queue", w.name)
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
close(w.quit)
}
func (w *WALWatcher) runWatcher() {
// The WAL dir may not exist when Prometheus first starts up.
for {
if _, err := os.Stat(w.walDir); os.IsNotExist(err) {
time.Sleep(time.Second)
} else {
break
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
}
}
nw, err := wal.New(nil, nil, w.walDir)
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
if err != nil {
level.Error(w.logger).Log("err", err)
return
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
}
first, last, err := nw.Segments()
if err != nil {
level.Error(w.logger).Log("err", err)
return
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
}
if last == -1 {
level.Error(w.logger).Log("err", err)
return
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
}
// Read series records in the current WAL and latest checkpoint, get the segment pointer back.
// TODO: callum, handle maintaining the WAL pointer somehow across apply configs?
segment, reader, err := w.readToEnd(w.walDir, first, last)
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
if err != nil {
level.Error(w.logger).Log("err", err)
return
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
}
w.currentSegment = last
w.currentSegmentMetric.Set(float64(w.currentSegment))
for {
level.Info(w.logger).Log("msg", "watching segment", "segment", w.currentSegment)
// On start, after reading the existing WAL for series records, we have a pointer to what is the latest segment.
// On subsequent calls to this function, currentSegment will have been incremented and we should open that segment.
err := w.watch(nw, reader)
segment.Close()
if err != nil {
level.Error(w.logger).Log("msg", "runWatcher is ending", "err", err)
return
}
w.currentSegment++
w.currentSegmentMetric.Set(float64(w.currentSegment))
segment, err = wal.OpenReadSegment(wal.SegmentName(w.walDir, w.currentSegment))
reader = wal.NewLiveReader(segment)
// TODO: callum, is this error really fatal?
if err != nil {
level.Error(w.logger).Log("err", err)
return
}
}
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
}
// When starting the WAL watcher, there is potentially an existing WAL. In that case, we
// should read to the end of the newest existing segment before reading new records that
// are written to it, storing data from series records along the way.
// Unfortunately this function is duplicates some of TSDB Head.Init().
func (w *WALWatcher) readToEnd(walDir string, firstSegment, lastSegment int) (*wal.Segment, *wal.LiveReader, error) {
// Backfill from the checkpoint first if it exists.
defer level.Debug(w.logger).Log("msg", "done reading existing WAL")
dir, startFrom, err := tsdb.LastCheckpoint(walDir)
if err != nil && err != tsdb.ErrNotFound {
return nil, nil, errors.Wrap(err, "find last checkpoint")
}
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
level.Debug(w.logger).Log("msg", "reading checkpoint", "dir", dir)
if err == nil {
w.lastCheckpoint = dir
err = w.readCheckpoint(dir)
if err != nil {
return nil, nil, err
}
startFrom++
}
// Backfill segments from the last checkpoint onwards if at least 2 segments exist.
if lastSegment > 0 {
for i := firstSegment; i < lastSegment; i++ {
seg, err := wal.OpenReadSegment(wal.SegmentName(walDir, i))
if err != nil {
return nil, nil, err
}
sz, _ := getSegmentSize(walDir, i)
w.readSeriesRecords(wal.NewLiveReader(seg), i, sz)
}
}
// We want to start the WAL Watcher from the end of the last segment on start,
// so we make sure to return the wal.Segment pointer
segment, err := wal.OpenReadSegment(wal.SegmentName(w.walDir, lastSegment))
if err != nil {
return nil, nil, err
}
r := wal.NewLiveReader(segment)
sz, _ := getSegmentSize(walDir, lastSegment)
w.readSeriesRecords(r, lastSegment, sz)
return segment, r, nil
}
// TODO: fix the exit logic for this function
// The stop param is used to stop at the end of the existing WAL on startup,
// since scraped samples may be written to the latest segment before we finish reading it.
func (w *WALWatcher) readSeriesRecords(r *wal.LiveReader, index int, stop int64) {
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
var (
dec tsdb.RecordDecoder
series []tsdb.RefSeries
samples []tsdb.RefSample
ret bool
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
)
for r.Next() && !isClosed(w.quit) {
series = series[:0]
rec := r.Record()
// If the timestamp is > start then we should Append this sample and exit readSeriesRecords,
// because this is the first sample written to the WAL after the WAL watcher was started.
typ := dec.Type(rec)
if typ == tsdb.RecordSamples {
samples, err := dec.Samples(rec, samples[:0])
if err != nil {
continue
}
for _, s := range samples {
if s.T > w.startTime {
w.writer.Append(samples)
ret = true
w.samplesSentPreTailing.Inc()
}
}
if ret {
level.Info(w.logger).Log("msg", "found a sample with a timestamp after the WAL watcher start")
level.Info(w.logger).Log("msg", "read all series records in segment/checkpoint", "index", index)
return
}
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
}
if typ != tsdb.RecordSeries {
continue
}
series, err := dec.Series(rec, nil)
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
if err != nil {
level.Error(log.With(w.logger)).Log("err", err)
break
}
w.writer.StoreSeries(series, index)
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
}
// Since we only call readSeriesRecords on fully written WAL segments or checkpoints,
// Error() will only return an error if something actually went wrong when reading
// a record, either it was invalid or it was only partially written to the WAL.
if err := r.Err(); err != nil {
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
level.Error(w.logger).Log("err", err)
return
}
// Ensure we read all of the bytes in the segment or checkpoint.
if r.TotalRead() >= stop {
level.Info(w.logger).Log("msg", "read all series records in segment/checkpoint", "index", index)
return
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
}
}
func (w *WALWatcher) watch(wl *wal.WAL, reader *wal.LiveReader) error {
readTicker := time.NewTicker(readPeriod)
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
defer readTicker.Stop()
checkpointTicker := time.NewTicker(checkpointPeriod)
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
defer checkpointTicker.Stop()
segmentTicker := time.NewTicker(segmentCheckPeriod)
defer segmentTicker.Stop()
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
for {
select {
case <-w.quit:
level.Info(w.logger).Log("msg", "quitting WAL watcher watch loop")
return errors.New("quit channel")
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
case <-checkpointTicker.C:
// Periodically check if there is a new checkpoint.
// As this is considered an optimisation, we ignore errors during
// checkpoint processing.
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
dir, _, err := tsdb.LastCheckpoint(w.walDir)
if err != nil && err != tsdb.ErrNotFound {
level.Error(w.logger).Log("msg", "error getting last checkpoint", "err", err)
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
continue
}
if dir == w.lastCheckpoint {
continue
}
level.Info(w.logger).Log("msg", "new checkpoint detected", "last", w.lastCheckpoint, "new", dir)
d, err := checkpointNum(dir)
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
if err != nil {
level.Error(w.logger).Log("msg", "error parsing checkpoint", "err", err)
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
continue
}
if d >= w.currentSegment {
level.Info(w.logger).Log("msg", "current segment is behind the checkpoint, skipping reading of checkpoint", "current", fmt.Sprintf("%08d", w.currentSegment), "checkpoint", dir)
continue
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
}
w.lastCheckpoint = dir
// This potentially takes a long time, should we run it in another go routine?
err = w.readCheckpoint(w.lastCheckpoint)
if err != nil {
level.Error(w.logger).Log("err", err)
}
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
case <-segmentTicker.C:
// check if new segments exist
_, last, err := wl.Segments()
if err != nil {
return errors.Wrap(err, "segments")
}
if last <= w.currentSegment {
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
continue
}
if err := w.readSegment(reader); err != nil {
// Ignore errors reading to end of segment, as we're going to move to
// next segment now.
level.Error(w.logger).Log("msg", "error reading to end of segment", "err", err)
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
}
level.Info(w.logger).Log("msg", "a new segment exists, we should start reading it", "current", fmt.Sprintf("%08d", w.currentSegment), "new", fmt.Sprintf("%08d", last))
return nil
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
case <-readTicker.C:
if err := w.readSegment(reader); err != nil {
level.Error(w.logger).Log("err", err)
return err
}
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
}
}
}
func (w *WALWatcher) readSegment(r *wal.LiveReader) error {
for r.Next() && !isClosed(w.quit) {
err := w.decodeRecord(r.Record())
// Intentionally skip over record decode errors.
if err != nil {
level.Error(w.logger).Log("err", err)
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
}
}
return r.Err()
}
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
func (w *WALWatcher) decodeRecord(rec []byte) error {
var (
dec tsdb.RecordDecoder
series []tsdb.RefSeries
samples []tsdb.RefSample
)
switch dec.Type(rec) {
case tsdb.RecordSeries:
series, err := dec.Series(rec, series[:0])
if err != nil {
w.recordDecodeFailsMetric.Inc()
return err
}
w.seriesReadMetric.Add(float64(len(series)))
w.writer.StoreSeries(series, w.currentSegment)
case tsdb.RecordSamples:
samples, err := dec.Samples(rec, samples[:0])
if err != nil {
w.recordDecodeFailsMetric.Inc()
return err
}
w.samplesReadMetric.Add(float64(len(samples)))
// Blocks until the sample is sent to all remote write endpoints or closed (because enqueue blocks).
w.writer.Append(samples)
case tsdb.RecordTombstones:
w.tombstonesReadMetric.Add(float64(len(samples)))
case tsdb.RecordInvalid:
w.invalidReadMetric.Add(float64(len(samples)))
return errors.New("invalid record")
default:
w.recordDecodeFailsMetric.Inc()
return errors.New("unknown TSDB record type")
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
}
return nil
}
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
// Read all the series records from a Checkpoint directory.
func (w *WALWatcher) readCheckpoint(checkpointDir string) error {
sr, err := wal.NewSegmentsReader(checkpointDir)
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
if err != nil {
return errors.Wrap(err, "open checkpoint")
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
}
defer sr.Close()
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
split := strings.Split(checkpointDir, ".")
if len(split) != 2 {
return errors.Errorf("checkpoint dir name is not in the right format: %s", checkpointDir)
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
}
i, err := strconv.Atoi(split[1])
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
if err != nil {
i = w.currentSegment - 1
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
}
size, err := getCheckpointSize(checkpointDir)
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
if err != nil {
level.Error(w.logger).Log("msg", "error getting checkpoint size", "checkpoint", checkpointDir)
return errors.Wrap(err, "get checkpoint size")
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
}
w.readSeriesRecords(wal.NewLiveReader(sr), i, size)
level.Debug(w.logger).Log("msg", "read series references from checkpoint", "checkpoint", checkpointDir)
w.writer.SeriesReset(i)
return nil
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
}
func checkpointNum(dir string) (int, error) {
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
// Checkpoint dir names are in the format checkpoint.000001
chunks := strings.Split(dir, ".")
if len(chunks) != 2 {
return 0, errors.Errorf("invalid checkpoint dir string: %s", dir)
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
}
result, err := strconv.Atoi(chunks[1])
if err != nil {
return 0, errors.Errorf("invalid checkpoint dir string: %s", dir)
}
return result, nil
Tail the TSDB WAL for remote_write This change switches the remote_write API to use the TSDB WAL. This should reduce memory usage and prevent sample loss when the remote end point is down. We use the new LiveReader from TSDB to tail WAL segments. Logic for finding the tracking segment is included in this PR. The WAL is tailed once for each remote_write endpoint specified. Reading from the segment is based on a ticker rather than relying on fsnotify write events, which were found to be complicated and unreliable in early prototypes. Enqueuing a sample for sending via remote_write can now block, to provide back pressure. Queues are still required to acheive parallelism and batching. We have updated the queue config based on new defaults for queue capacity and pending samples values - much smaller values are now possible. The remote_write resharding code has been updated to prevent deadlocks, and extra tests have been added for these cases. As part of this change, we attempt to guarantee that samples are not lost; however this initial version doesn't guarantee this across Prometheus restarts or non-retryable errors from the remote end (eg 400s). This changes also includes the following optimisations: - only marshal the proto request once, not once per retry - maintain a single copy of the labels for given series to reduce GC pressure Other minor tweaks: - only reshard if we've also successfully sent recently - add pending samples, latest sent timestamp, WAL events processed metrics Co-authored-by: Chris Marchbanks <csmarchbanks.com> (initial prototype) Co-authored-by: Tom Wilkie <tom.wilkie@gmail.com> (sharding changes) Signed-off-by: Callum Styan <callumstyan@gmail.com>
2018-09-07 14:26:04 -07:00
}
func getCheckpointSize(dir string) (int64, error) {
i := int64(0)
segs, err := fileutil.ReadDir(dir)
if err != nil {
return 0, err
}
for _, fn := range segs {
num, err := strconv.Atoi(fn)
if err != nil {
return i, err
}
sz, err := getSegmentSize(dir, num)
if err != nil {
return i, err
}
i += sz
}
return i, nil
}
// Get size of segment.
func getSegmentSize(dir string, index int) (int64, error) {
i := int64(-1)
fi, err := os.Stat(wal.SegmentName(dir, index))
if err == nil {
i = fi.Size()
}
return i, err
}
func isClosed(c chan struct{}) bool {
select {
case <-c:
return true
default:
return false
}
}