The buffers we allocated were escaping to the heap, resulting in large
memory usage spikes during startup and checkpointing in Prometheus.
This attaches the buffer to the reader object to prevent this.
Signed-off-by: Fabian Reinartz <freinartz@google.com>
We assume in multiple places that the block list held by DB
has blocks sequential by time.
A regression caused us to hold them ordered by ULID, i.e. by creation
time instead.
Signed-off-by: Fabian Reinartz <freinartz@google.com>
Create checkpoints from a sequence of WAL segments while filtering
out obsolete data. The checkpoint format is again a sequence of WAL
segments, which allows us to reuse the serialization format and
implementation.
Signed-off-by: Fabian Reinartz <freinartz@google.com>
Allow to repair the WAL based on the error returned by a reader
during a full scan over all records.
Signed-off-by: Fabian Reinartz <freinartz@google.com>
This adds a new WAL that's agnostic to the actual record contents.
It's much simpler and should be more resilient than the existing one.
Signed-off-by: Fabian Reinartz <freinartz@google.com>
Due to the way blocks used to overlap by 1 millisecond (see #347), when
requesting a 2-hour interval starting at `blocks[1].MinTime`, the
`Querier` would consider three blocks: `blocks[0]`, `blocks[1]` and
`blocks[2]`, because `blocks[0].MaxTime` and `blocks[2].MinTime` were in
that interval.
However, if the blocks don't overlap, only two blocks should be
returned: `blocks[1]` and `blocks[2]`. This test ensures that it's
indeed the case.
Signed-off-by: Benoît Knecht <benoit.knecht@fsfe.org>
Blocks are half-open intervals [a, b), while all other intervals
(chunks, head, ...) are closed intervals [a, b].
Make that distinction explicit by defining `OverlapsClosedInterval()`
methods for blocks and chunks, and using them in place of the more
generic `intervalOverlap()` function.
This change also fixes `db.Querier()` and `db.Delete()`, which could
previously return one extraneous block at the end of the specified
interval.
Signed-off-by: Benoît Knecht <benoit.knecht@fsfe.org>
Block intervals are bound by `block.MinTime`, `block.MaxTime`, but they
define a half-open interval: `[block.MinTime, block.MaxTime).
However, when deciding if a chunk was part of a block or not, the
`intervalOverlap()` function would consider both the chunk and the block
intervals as being closed.
Rather than modify the login in `intervalOverlap()`, we explicitly
remove the last value from the interval when reading from head to
persist blocks.
Signed-off-by: Benoît Knecht <benoit.knecht@fsfe.org>
Due to a regression blocks were no longer ordered by time before
being stored in memory. This made data intermittently become unavailable
for queries.
Signed-off-by: Fabian Reinartz <freinartz@google.com>
