prometheus/storage/remote/queue_manager_test.go

805 lines
23 KiB
Go
Raw Normal View History

2017-05-10 02:44:13 -07:00
// Copyright 2013 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 (
"context"
2017-05-10 02:44:13 -07:00
"fmt"
"io/ioutil"
"math"
"net/url"
"os"
"sort"
"strconv"
"strings"
2017-05-10 02:44:13 -07:00
"sync"
"testing"
"time"
"github.com/go-kit/kit/log"
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
"github.com/gogo/protobuf/proto"
"github.com/golang/snappy"
"github.com/prometheus/client_golang/prometheus"
client_testutil "github.com/prometheus/client_golang/prometheus/testutil"
common_config "github.com/prometheus/common/config"
2017-05-10 02:44:13 -07:00
"github.com/prometheus/common/model"
"github.com/stretchr/testify/require"
"go.uber.org/atomic"
"github.com/prometheus/prometheus/config"
"github.com/prometheus/prometheus/pkg/labels"
"github.com/prometheus/prometheus/pkg/textparse"
"github.com/prometheus/prometheus/pkg/timestamp"
2017-10-23 13:28:17 -07:00
"github.com/prometheus/prometheus/prompb"
"github.com/prometheus/prometheus/scrape"
"github.com/prometheus/prometheus/tsdb/record"
2017-05-10 02:44:13 -07:00
)
const defaultFlushDeadline = 1 * time.Minute
func newHighestTimestampMetric() *maxTimestamp {
return &maxTimestamp{
Gauge: prometheus.NewGauge(prometheus.GaugeOpts{
Namespace: namespace,
Subsystem: subsystem,
Name: "highest_timestamp_in_seconds",
Help: "Highest timestamp that has come into the remote storage via the Appender interface, in seconds since epoch.",
}),
}
}
2017-05-10 02:44:13 -07:00
func TestSampleDelivery(t *testing.T) {
// Let's create an even number of send batches so we don't run into the
// batch timeout case.
n := config.DefaultQueueConfig.MaxSamplesPerSend * 2
samples, series := createTimeseries(n, n)
2017-05-10 02:44:13 -07:00
c := NewTestWriteClient()
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
c.expectSamples(samples[:len(samples)/2], series)
2017-05-10 02:44:13 -07:00
queueConfig := config.DefaultQueueConfig
queueConfig.BatchSendDeadline = model.Duration(100 * time.Millisecond)
queueConfig.MaxShards = 1
queueConfig.Capacity = len(samples)
queueConfig.MaxSamplesPerSend = len(samples) / 2
dir, err := ioutil.TempDir("", "TestSampleDeliver")
require.NoError(t, err)
defer func() {
require.NoError(t, os.RemoveAll(dir))
}()
s := NewStorage(nil, nil, nil, dir, defaultFlushDeadline, nil)
defer s.Close()
2017-05-10 02:44:13 -07:00
writeConfig := config.DefaultRemoteWriteConfig
conf := &config.Config{
GlobalConfig: config.DefaultGlobalConfig,
RemoteWriteConfigs: []*config.RemoteWriteConfig{
&writeConfig,
},
}
// We need to set URL's so that metric creation doesn't panic.
writeConfig.URL = &common_config.URL{
URL: &url.URL{
Host: "http://test-storage.com",
},
}
writeConfig.QueueConfig = queueConfig
require.NoError(t, s.ApplyConfig(conf))
hash, err := toHash(writeConfig)
require.NoError(t, err)
qm := s.rws.queues[hash]
qm.SetClient(c)
qm.StoreSeries(series, 0)
2017-05-10 02:44:13 -07:00
qm.Append(samples[:len(samples)/2])
2017-05-10 02:44:13 -07:00
c.waitForExpectedSamples(t)
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
c.expectSamples(samples[len(samples)/2:], series)
qm.Append(samples[len(samples)/2:])
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
c.waitForExpectedSamples(t)
2017-05-10 02:44:13 -07:00
}
func TestMetadataDelivery(t *testing.T) {
c := NewTestWriteClient()
dir, err := ioutil.TempDir("", "TestMetadataDelivery")
require.NoError(t, err)
defer os.RemoveAll(dir)
cfg := config.DefaultQueueConfig
mcfg := config.DefaultMetadataConfig
metrics := newQueueManagerMetrics(nil, "", "")
m := NewQueueManager(metrics, nil, nil, nil, dir, newEWMARate(ewmaWeight, shardUpdateDuration), cfg, mcfg, nil, nil, c, defaultFlushDeadline, newPool(), newHighestTimestampMetric(), nil)
m.Start()
defer m.Stop()
m.AppendMetadata(context.Background(), []scrape.MetricMetadata{
scrape.MetricMetadata{
Metric: "prometheus_remote_storage_sent_metadata_bytes_total",
Type: textparse.MetricTypeCounter,
Help: "a nice help text",
Unit: "",
},
})
require.Equal(t, len(c.receivedMetadata), 1)
}
2018-03-12 08:35:43 -07:00
func TestSampleDeliveryTimeout(t *testing.T) {
2018-03-12 09:48:51 -07:00
// Let's send one less sample than batch size, and wait the timeout duration
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
n := 9
samples, series := createTimeseries(n, n)
c := NewTestWriteClient()
2018-03-12 08:35:43 -07:00
cfg := config.DefaultQueueConfig
mcfg := config.DefaultMetadataConfig
2018-03-12 08:35:43 -07:00
cfg.MaxShards = 1
cfg.BatchSendDeadline = model.Duration(100 * time.Millisecond)
dir, err := ioutil.TempDir("", "TestSampleDeliveryTimeout")
require.NoError(t, err)
defer func() {
require.NoError(t, os.RemoveAll(dir))
}()
metrics := newQueueManagerMetrics(nil, "", "")
m := NewQueueManager(metrics, nil, nil, nil, dir, newEWMARate(ewmaWeight, shardUpdateDuration), cfg, mcfg, nil, nil, c, defaultFlushDeadline, newPool(), newHighestTimestampMetric(), nil)
m.StoreSeries(series, 0)
2018-03-12 08:35:43 -07:00
m.Start()
defer m.Stop()
// Send the samples twice, waiting for the samples in the meantime.
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
c.expectSamples(samples, series)
m.Append(samples)
2018-03-12 08:35:43 -07:00
c.waitForExpectedSamples(t)
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
c.expectSamples(samples, series)
m.Append(samples)
2018-03-12 08:35:43 -07:00
c.waitForExpectedSamples(t)
}
2017-05-10 02:44:13 -07:00
func TestSampleDeliveryOrder(t *testing.T) {
ts := 10
n := config.DefaultQueueConfig.MaxSamplesPerSend * ts
samples := make([]record.RefSample, 0, n)
series := make([]record.RefSeries, 0, n)
2017-05-10 02:44:13 -07:00
for i := 0; i < n; i++ {
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
name := fmt.Sprintf("test_metric_%d", i%ts)
samples = append(samples, record.RefSample{
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
Ref: uint64(i),
T: int64(i),
V: float64(i),
})
series = append(series, record.RefSeries{
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
Ref: uint64(i),
Labels: labels.Labels{labels.Label{Name: "__name__", Value: name}},
2017-05-10 02:44:13 -07:00
})
}
c := NewTestWriteClient()
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
c.expectSamples(samples, series)
dir, err := ioutil.TempDir("", "TestSampleDeliveryOrder")
require.NoError(t, err)
defer func() {
require.NoError(t, os.RemoveAll(dir))
}()
cfg := config.DefaultQueueConfig
mcfg := config.DefaultMetadataConfig
metrics := newQueueManagerMetrics(nil, "", "")
m := NewQueueManager(metrics, nil, nil, nil, dir, newEWMARate(ewmaWeight, shardUpdateDuration), cfg, mcfg, nil, nil, c, defaultFlushDeadline, newPool(), newHighestTimestampMetric(), nil)
m.StoreSeries(series, 0)
2017-05-10 02:44:13 -07:00
m.Start()
defer m.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
// These should be received by the client.
m.Append(samples)
2017-05-10 02:44:13 -07:00
c.waitForExpectedSamples(t)
}
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 TestShutdown(t *testing.T) {
deadline := 1 * time.Second
c := NewTestBlockedWriteClient()
2017-05-10 02:44:13 -07:00
dir, err := ioutil.TempDir("", "TestShutdown")
require.NoError(t, err)
defer func() {
require.NoError(t, os.RemoveAll(dir))
}()
cfg := config.DefaultQueueConfig
mcfg := config.DefaultMetadataConfig
metrics := newQueueManagerMetrics(nil, "", "")
m := NewQueueManager(metrics, nil, nil, nil, dir, newEWMARate(ewmaWeight, shardUpdateDuration), cfg, mcfg, nil, nil, c, deadline, newPool(), newHighestTimestampMetric(), nil)
n := 2 * config.DefaultQueueConfig.MaxSamplesPerSend
samples, series := createTimeseries(n, n)
m.StoreSeries(series, 0)
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
m.Start()
// Append blocks to guarantee delivery, so we do it in the background.
go func() {
m.Append(samples)
}()
time.Sleep(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
// Test to ensure that Stop doesn't block.
start := time.Now()
m.Stop()
// The samples will never be delivered, so duration should
// be at least equal to deadline, otherwise the flush deadline
// was not respected.
duration := time.Since(start)
if duration > time.Duration(deadline+(deadline/10)) {
t.Errorf("Took too long to shutdown: %s > %s", duration, deadline)
}
if duration < time.Duration(deadline) {
t.Errorf("Shutdown occurred before flush deadline: %s < %s", duration, deadline)
2017-05-10 02:44:13 -07:00
}
}
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 TestSeriesReset(t *testing.T) {
c := NewTestBlockedWriteClient()
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
deadline := 5 * time.Second
numSegments := 4
numSeries := 25
dir, err := ioutil.TempDir("", "TestSeriesReset")
require.NoError(t, err)
defer func() {
require.NoError(t, os.RemoveAll(dir))
}()
cfg := config.DefaultQueueConfig
mcfg := config.DefaultMetadataConfig
metrics := newQueueManagerMetrics(nil, "", "")
m := NewQueueManager(metrics, nil, nil, nil, dir, newEWMARate(ewmaWeight, shardUpdateDuration), cfg, mcfg, nil, nil, c, deadline, newPool(), newHighestTimestampMetric(), 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
for i := 0; i < numSegments; i++ {
series := []record.RefSeries{}
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 j := 0; j < numSeries; j++ {
series = append(series, record.RefSeries{Ref: uint64((i * 100) + j), Labels: labels.Labels{{Name: "a", Value: "a"}}})
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
}
m.StoreSeries(series, i)
}
require.Equal(t, numSegments*numSeries, len(m.seriesLabels))
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
m.SeriesReset(2)
require.Equal(t, numSegments*numSeries/2, len(m.seriesLabels))
2017-05-10 02:44:13 -07:00
}
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 TestReshard(t *testing.T) {
size := 10 // Make bigger to find more races.
nSeries := 6
nSamples := config.DefaultQueueConfig.Capacity * size
samples, series := createTimeseries(nSamples, nSeries)
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
c := NewTestWriteClient()
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
c.expectSamples(samples, series)
cfg := config.DefaultQueueConfig
mcfg := config.DefaultMetadataConfig
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
cfg.MaxShards = 1
dir, err := ioutil.TempDir("", "TestReshard")
require.NoError(t, err)
defer func() {
require.NoError(t, os.RemoveAll(dir))
}()
metrics := newQueueManagerMetrics(nil, "", "")
m := NewQueueManager(metrics, nil, nil, nil, dir, newEWMARate(ewmaWeight, shardUpdateDuration), cfg, mcfg, nil, nil, c, defaultFlushDeadline, newPool(), newHighestTimestampMetric(), nil)
m.StoreSeries(series, 0)
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
m.Start()
defer m.Stop()
go func() {
for i := 0; i < len(samples); i += config.DefaultQueueConfig.Capacity {
sent := m.Append(samples[i : i+config.DefaultQueueConfig.Capacity])
require.True(t, sent, "samples not sent")
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
time.Sleep(100 * time.Millisecond)
}
}()
for i := 1; i < len(samples)/config.DefaultQueueConfig.Capacity; i++ {
m.shards.stop()
m.shards.start(i)
time.Sleep(100 * time.Millisecond)
}
c.waitForExpectedSamples(t)
2017-05-10 02:44:13 -07:00
}
func TestReshardRaceWithStop(t *testing.T) {
c := NewTestWriteClient()
var m *QueueManager
h := sync.Mutex{}
h.Lock()
cfg := config.DefaultQueueConfig
mcfg := config.DefaultMetadataConfig
go func() {
for {
metrics := newQueueManagerMetrics(nil, "", "")
m = NewQueueManager(metrics, nil, nil, nil, "", newEWMARate(ewmaWeight, shardUpdateDuration), cfg, mcfg, nil, nil, c, defaultFlushDeadline, newPool(), newHighestTimestampMetric(), nil)
m.Start()
h.Unlock()
h.Lock()
m.Stop()
}
}()
for i := 1; i < 100; i++ {
h.Lock()
m.reshardChan <- i
h.Unlock()
}
}
func TestReleaseNoninternedString(t *testing.T) {
cfg := config.DefaultQueueConfig
mcfg := config.DefaultMetadataConfig
metrics := newQueueManagerMetrics(nil, "", "")
c := NewTestWriteClient()
m := NewQueueManager(metrics, nil, nil, nil, "", newEWMARate(ewmaWeight, shardUpdateDuration), cfg, mcfg, nil, nil, c, defaultFlushDeadline, newPool(), newHighestTimestampMetric(), nil)
m.Start()
for i := 1; i < 1000; i++ {
m.StoreSeries([]record.RefSeries{
{
Ref: uint64(i),
Labels: labels.Labels{
labels.Label{
Name: "asdf",
Value: fmt.Sprintf("%d", i),
},
},
},
}, 0)
m.SeriesReset(1)
}
metric := client_testutil.ToFloat64(noReferenceReleases)
require.Equal(t, 0.0, metric, "expected there to be no calls to release for strings that were not already interned: %d", int(metric))
}
func TestShouldReshard(t *testing.T) {
type testcase struct {
startingShards int
samplesIn, samplesOut, lastSendTimestamp int64
expectedToReshard bool
}
cases := []testcase{
{
// Resharding shouldn't take place if the last successful send was > batch send deadline*2 seconds ago.
startingShards: 10,
samplesIn: 1000,
samplesOut: 10,
lastSendTimestamp: time.Now().Unix() - int64(3*time.Duration(config.DefaultQueueConfig.BatchSendDeadline)/time.Second),
expectedToReshard: false,
},
{
startingShards: 5,
samplesIn: 1000,
samplesOut: 10,
lastSendTimestamp: time.Now().Unix(),
expectedToReshard: true,
},
}
cfg := config.DefaultQueueConfig
mcfg := config.DefaultMetadataConfig
for _, c := range cases {
metrics := newQueueManagerMetrics(nil, "", "")
client := NewTestWriteClient()
m := NewQueueManager(metrics, nil, nil, nil, "", newEWMARate(ewmaWeight, shardUpdateDuration), cfg, mcfg, nil, nil, client, defaultFlushDeadline, newPool(), newHighestTimestampMetric(), nil)
m.numShards = c.startingShards
m.samplesIn.incr(c.samplesIn)
m.samplesOut.incr(c.samplesOut)
m.lastSendTimestamp.Store(c.lastSendTimestamp)
m.Start()
desiredShards := m.calculateDesiredShards()
shouldReshard := m.shouldReshard(desiredShards)
m.Stop()
require.Equal(t, c.expectedToReshard, shouldReshard)
}
}
func createTimeseries(numSamples, numSeries int) ([]record.RefSample, []record.RefSeries) {
samples := make([]record.RefSample, 0, numSamples)
series := make([]record.RefSeries, 0, numSeries)
for i := 0; i < numSeries; i++ {
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
name := fmt.Sprintf("test_metric_%d", i)
for j := 0; j < numSamples; j++ {
samples = append(samples, record.RefSample{
Ref: uint64(i),
T: int64(j),
V: float64(i),
})
}
series = append(series, record.RefSeries{
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
Ref: uint64(i),
Labels: labels.Labels{{Name: "__name__", Value: 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 samples, series
2017-05-10 02:44:13 -07:00
}
func getSeriesNameFromRef(r record.RefSeries) 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
for _, l := range r.Labels {
if l.Name == "__name__" {
return l.Value
}
}
return ""
2017-05-10 02:44:13 -07:00
}
type TestWriteClient struct {
receivedSamples map[string][]prompb.Sample
expectedSamples map[string][]prompb.Sample
receivedMetadata map[string][]prompb.MetricMetadata
withWaitGroup bool
wg sync.WaitGroup
mtx sync.Mutex
buf []byte
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
}
2017-05-10 02:44:13 -07:00
func NewTestWriteClient() *TestWriteClient {
return &TestWriteClient{
withWaitGroup: true,
receivedSamples: map[string][]prompb.Sample{},
expectedSamples: map[string][]prompb.Sample{},
receivedMetadata: map[string][]prompb.MetricMetadata{},
2017-05-10 02:44:13 -07:00
}
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
}
2017-05-10 02:44:13 -07:00
func (c *TestWriteClient) expectSamples(ss []record.RefSample, series []record.RefSeries) {
if !c.withWaitGroup {
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
c.mtx.Lock()
defer c.mtx.Unlock()
2017-05-10 02:44:13 -07:00
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
c.expectedSamples = map[string][]prompb.Sample{}
c.receivedSamples = map[string][]prompb.Sample{}
2017-05-10 02:44:13 -07:00
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 _, s := range ss {
seriesName := getSeriesNameFromRef(series[s.Ref])
c.expectedSamples[seriesName] = append(c.expectedSamples[seriesName], prompb.Sample{
Timestamp: s.T,
Value: s.V,
})
}
c.wg.Add(len(ss))
}
2017-05-10 02:44:13 -07:00
func (c *TestWriteClient) waitForExpectedSamples(tb testing.TB) {
if !c.withWaitGroup {
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
c.wg.Wait()
c.mtx.Lock()
defer c.mtx.Unlock()
for ts, expectedSamples := range c.expectedSamples {
require.Equal(tb, expectedSamples, c.receivedSamples[ts], ts)
2017-05-10 02:44:13 -07:00
}
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
}
2017-05-10 02:44:13 -07:00
func (c *TestWriteClient) expectSampleCount(numSamples int) {
if !c.withWaitGroup {
return
}
c.mtx.Lock()
defer c.mtx.Unlock()
c.wg.Add(numSamples)
}
func (c *TestWriteClient) waitForExpectedSampleCount() {
if !c.withWaitGroup {
return
}
c.wg.Wait()
}
func (c *TestWriteClient) Store(_ context.Context, req []byte) 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
c.mtx.Lock()
defer c.mtx.Unlock()
// nil buffers are ok for snappy, ignore cast error.
if c.buf != nil {
c.buf = c.buf[:cap(c.buf)]
}
reqBuf, err := snappy.Decode(c.buf, req)
c.buf = reqBuf
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 err
2017-05-10 02:44:13 -07:00
}
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 reqProto prompb.WriteRequest
if err := proto.Unmarshal(reqBuf, &reqProto); err != nil {
return err
2017-05-10 02:44:13 -07:00
}
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
count := 0
for _, ts := range reqProto.Timeseries {
var seriesName string
labels := labelProtosToLabels(ts.Labels)
for _, label := range labels {
if label.Name == "__name__" {
seriesName = label.Value
}
}
for _, sample := range ts.Samples {
count++
c.receivedSamples[seriesName] = append(c.receivedSamples[seriesName], sample)
}
2017-05-10 02:44:13 -07:00
}
if c.withWaitGroup {
c.wg.Add(-count)
}
for _, m := range reqProto.Metadata {
c.receivedMetadata[m.MetricFamilyName] = append(c.receivedMetadata[m.MetricFamilyName], m)
}
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
2017-05-10 02:44:13 -07:00
}
func (c *TestWriteClient) Name() string {
return "testwriteclient"
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 (c *TestWriteClient) Endpoint() string {
return "http://test-remote.com/1234"
}
// TestBlockingWriteClient is a queue_manager WriteClient which will block
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
// on any calls to Store(), until the request's Context is cancelled, at which
// point the `numCalls` property will contain a count of how many times Store()
// was called.
type TestBlockingWriteClient struct {
numCalls atomic.Uint64
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 NewTestBlockedWriteClient() *TestBlockingWriteClient {
return &TestBlockingWriteClient{}
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 (c *TestBlockingWriteClient) Store(ctx context.Context, _ []byte) error {
c.numCalls.Inc()
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
<-ctx.Done()
return nil
}
func (c *TestBlockingWriteClient) NumCalls() uint64 {
return c.numCalls.Load()
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 (c *TestBlockingWriteClient) Name() string {
return "testblockingwriteclient"
}
func (c *TestBlockingWriteClient) Endpoint() string {
return "http://test-remote-blocking.com/1234"
}
func BenchmarkSampleDelivery(b *testing.B) {
// Let's create an even number of send batches so we don't run into the
// batch timeout case.
n := config.DefaultQueueConfig.MaxSamplesPerSend * 10
samples, series := createTimeseries(n, n)
c := NewTestWriteClient()
cfg := config.DefaultQueueConfig
mcfg := config.DefaultMetadataConfig
cfg.BatchSendDeadline = model.Duration(100 * time.Millisecond)
cfg.MaxShards = 1
dir, err := ioutil.TempDir("", "BenchmarkSampleDelivery")
require.NoError(b, err)
defer os.RemoveAll(dir)
metrics := newQueueManagerMetrics(nil, "", "")
m := NewQueueManager(metrics, nil, nil, nil, dir, newEWMARate(ewmaWeight, shardUpdateDuration), cfg, mcfg, nil, nil, c, defaultFlushDeadline, newPool(), newHighestTimestampMetric(), nil)
m.StoreSeries(series, 0)
// These should be received by the client.
m.Start()
defer m.Stop()
b.ResetTimer()
for i := 0; i < b.N; i++ {
c.expectSampleCount(len(samples))
m.Append(samples)
c.waitForExpectedSampleCount()
}
// Do not include shutdown
b.StopTimer()
}
func BenchmarkStartup(b *testing.B) {
dir := os.Getenv("WALDIR")
if dir == "" {
return
}
// Find the second largest segment; we will replay up to this.
// (Second largest as WALWatcher will start tailing the largest).
dirents, err := ioutil.ReadDir(dir)
require.NoError(b, err)
var segments []int
for _, dirent := range dirents {
if i, err := strconv.Atoi(dirent.Name()); err != nil {
segments = append(segments, i)
}
}
sort.Ints(segments)
logger := log.NewLogfmtLogger(log.NewSyncWriter(os.Stdout))
logger = log.With(logger, "caller", log.DefaultCaller)
cfg := config.DefaultQueueConfig
mcfg := config.DefaultMetadataConfig
for n := 0; n < b.N; n++ {
metrics := newQueueManagerMetrics(nil, "", "")
c := NewTestBlockedWriteClient()
m := NewQueueManager(metrics, nil, nil, logger, dir,
newEWMARate(ewmaWeight, shardUpdateDuration),
cfg, mcfg, nil, nil, c, 1*time.Minute, newPool(), newHighestTimestampMetric(), nil)
m.watcher.SetStartTime(timestamp.Time(math.MaxInt64))
m.watcher.MaxSegment = segments[len(segments)-2]
err := m.watcher.Run()
require.NoError(b, err)
}
}
func TestProcessExternalLabels(t *testing.T) {
for _, tc := range []struct {
labels labels.Labels
externalLabels labels.Labels
expected labels.Labels
}{
// Test adding labels at the end.
{
labels: labels.Labels{{Name: "a", Value: "b"}},
externalLabels: labels.Labels{{Name: "c", Value: "d"}},
expected: labels.Labels{{Name: "a", Value: "b"}, {Name: "c", Value: "d"}},
},
// Test adding labels at the beginning.
{
labels: labels.Labels{{Name: "c", Value: "d"}},
externalLabels: labels.Labels{{Name: "a", Value: "b"}},
expected: labels.Labels{{Name: "a", Value: "b"}, {Name: "c", Value: "d"}},
},
// Test we don't override existing labels.
{
labels: labels.Labels{{Name: "a", Value: "b"}},
externalLabels: labels.Labels{{Name: "a", Value: "c"}},
expected: labels.Labels{{Name: "a", Value: "b"}},
},
} {
require.Equal(t, tc.expected, processExternalLabels(tc.labels, tc.externalLabels))
}
}
func TestCalculateDesiredShards(t *testing.T) {
c := NewTestWriteClient()
cfg := config.DefaultQueueConfig
mcfg := config.DefaultMetadataConfig
dir, err := ioutil.TempDir("", "TestCalculateDesiredShards")
require.NoError(t, err)
defer func() {
require.NoError(t, os.RemoveAll(dir))
}()
metrics := newQueueManagerMetrics(nil, "", "")
samplesIn := newEWMARate(ewmaWeight, shardUpdateDuration)
m := NewQueueManager(metrics, nil, nil, nil, dir, samplesIn, cfg, mcfg, nil, nil, c, defaultFlushDeadline, newPool(), newHighestTimestampMetric(), nil)
// Need to start the queue manager so the proper metrics are initialized.
// However we can stop it right away since we don't need to do any actual
// processing.
m.Start()
m.Stop()
inputRate := int64(50000)
var pendingSamples int64
// Two minute startup, no samples are sent.
startedAt := time.Now().Add(-2 * time.Minute)
// helper function for adding samples.
addSamples := func(s int64, ts time.Duration) {
pendingSamples += s
samplesIn.incr(s)
samplesIn.tick()
m.highestRecvTimestamp.Set(float64(startedAt.Add(ts).Unix()))
}
// helper function for sending samples.
sendSamples := func(s int64, ts time.Duration) {
pendingSamples -= s
m.samplesOut.incr(s)
m.samplesOutDuration.incr(int64(m.numShards) * int64(shardUpdateDuration))
// highest sent is how far back pending samples would be at our input rate.
highestSent := startedAt.Add(ts - time.Duration(pendingSamples/inputRate)*time.Second)
m.metrics.highestSentTimestamp.Set(float64(highestSent.Unix()))
m.lastSendTimestamp.Store(time.Now().Unix())
}
ts := time.Duration(0)
for ; ts < 120*time.Second; ts += shardUpdateDuration {
addSamples(inputRate*int64(shardUpdateDuration/time.Second), ts)
m.numShards = m.calculateDesiredShards()
require.Equal(t, 1, m.numShards)
}
// Assume 100ms per request, or 10 requests per second per shard.
// Shard calculation should never drop below barely keeping up.
minShards := int(inputRate) / cfg.MaxSamplesPerSend / 10
// This test should never go above 200 shards, that would be more resources than needed.
maxShards := 200
for ; ts < 15*time.Minute; ts += shardUpdateDuration {
sin := inputRate * int64(shardUpdateDuration/time.Second)
addSamples(sin, ts)
sout := int64(m.numShards*cfg.MaxSamplesPerSend) * int64(shardUpdateDuration/(100*time.Millisecond))
// You can't send samples that don't exist so cap at the number of pending samples.
if sout > pendingSamples {
sout = pendingSamples
}
sendSamples(sout, ts)
t.Log("desiredShards", m.numShards, "pendingSamples", pendingSamples)
m.numShards = m.calculateDesiredShards()
require.GreaterOrEqual(t, m.numShards, minShards, "Shards are too low. desiredShards=%d, minShards=%d, t_seconds=%d", m.numShards, minShards, ts/time.Second)
require.LessOrEqual(t, m.numShards, maxShards, "Shards are too high. desiredShards=%d, maxShards=%d, t_seconds=%d", m.numShards, maxShards, ts/time.Second)
}
require.Equal(t, int64(0), pendingSamples, "Remote write never caught up, there are still %d pending samples.", pendingSamples)
}
func TestQueueManagerMetrics(t *testing.T) {
reg := prometheus.NewPedanticRegistry()
metrics := newQueueManagerMetrics(reg, "name", "http://localhost:1234")
// Make sure metrics pass linting.
problems, err := client_testutil.GatherAndLint(reg)
require.NoError(t, err)
require.Equal(t, 0, len(problems), "Metric linting problems detected: %v", problems)
// Make sure all metrics were unregistered. A failure here means you need
// unregister a metric in `queueManagerMetrics.unregister()`.
metrics.unregister()
err = client_testutil.GatherAndCompare(reg, strings.NewReader(""))
require.NoError(t, err)
}