* Callbacks for lifecycle of series in TSDB
Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in>
* Add more comments
Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in>
When appending to the head and a chunk is full it is flushed to the disk and m-mapped (memory mapped) to free up memory
Prom startup now happens in these stages
- Iterate the m-maped chunks from disk and keep a map of series reference to its slice of mmapped chunks.
- Iterate the WAL as usual. Whenever we create a new series, look for it's mmapped chunks in the map created before and add it to that series.
If a head chunk is corrupted the currpted one and all chunks after that are deleted and the data after the corruption is recovered from the existing WAL which means that a corruption in m-mapped files results in NO data loss.
[Mmaped chunks format](https://github.com/prometheus/prometheus/blob/master/tsdb/docs/format/head_chunks.md) - main difference is that the chunk for mmaping now also includes series reference because there is no index for mapping series to chunks.
[The block chunks](https://github.com/prometheus/prometheus/blob/master/tsdb/docs/format/chunks.md) are accessed from the index which includes the offsets for the chunks in the chunks file - example - chunks of series ID have offsets 200, 500 etc in the chunk files.
In case of mmaped chunks, the offsets are stored in memory and accessed from that. During WAL replay, these offsets are restored by iterating all m-mapped chunks as stated above by matching the series id present in the chunk header and offset of that chunk in that file.
**Prombench results**
_WAL Replay_
1h Wal reply time
30% less wal reply time - 4m31 vs 3m36
2h Wal reply time
20% less wal reply time - 8m16 vs 7m
_Memory During WAL Replay_
High Churn:
10-15% less RAM - 32gb vs 28gb
20% less RAM after compaction 34gb vs 27gb
No Churn:
20-30% less RAM - 23gb vs 18gb
40% less RAM after compaction 32.5gb vs 20gb
Screenshots are in [this comment](https://github.com/prometheus/prometheus/pull/6679#issuecomment-621678932)
Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in>
Prior to this commit we could have situations where we are creating an
appenderId but never creating an appender to go with it, therefore
blocking the low watermak.
Signed-off-by: Julien Pivotto <roidelapluie@inuits.eu>
Previously we were keeping up to around 6 hours of WAL around by
removing 1/3 every hours. This was excessive, so switch to removing 2/3
which will up to around 3 hours of WAL around.
This will roughly halve the size of the WAL and halve startup time for
those who are I/O bound. This may increase the checkpoint size for
those with certain churn patterns, but by much less than we're saving
from the segments.
Signed-off-by: Brian Brazil <brian.brazil@robustperception.io>
time.Unix attaches the local timezone, which can then
leak out (e.g. in the alert json). While this is harmless,
we should be consistent.
Signed-off-by: Brian Brazil <brian.brazil@robustperception.io>
* storage: Added Chunks{Queryable/Querier/SeriesSet/Series/Iteratable. Added generic Merge{SeriesSet/Querier} implementation.
## Rationales:
In many places (e.g. chunk Remote read, Thanos Receive fetching chunk from TSDB), we operate on encoded chunks not samples.
This means that we unnecessary decode/encode, wasting CPU, time and memory.
This PR adds chunk iterator interfaces and makes the merge code to be reused between both seriesSets
I will make the use of it in following PR inside tsdb itself. For now fanout implements it and mergers.
All merges now also allows passing series mergers. This opens doors for custom deduplications other than TSDB vertical ones (e.g. offline one we have in Thanos).
## Changes
* Added Chunk versions of all iterating methods. It all starts in Querier/ChunkQuerier. The plan is that
Storage will implement both chunked and samples.
* Added Seek to chunks.Iterator interface for iterating over chunks.
* NewMergeChunkQuerier was added; Both this and NewMergeQuerier are now using generigMergeQuerier to share the code. Generic code was added.
* Improved tests.
* Added some TODO for further simplifications in next PRs.
Signed-off-by: Bartlomiej Plotka <bwplotka@gmail.com>
* Addressed Brian's comments.
Signed-off-by: Bartlomiej Plotka <bwplotka@gmail.com>
* Moved s/Labeled/SeriesLabels as per Krasi suggestion.
Signed-off-by: Bartlomiej Plotka <bwplotka@gmail.com>
* Addressed Krasi's comments.
Signed-off-by: Bartlomiej Plotka <bwplotka@gmail.com>
* Second iteration of Krasi comments.
Signed-off-by: Bartlomiej Plotka <bwplotka@gmail.com>
* Another round of comments.
Signed-off-by: Bartlomiej Plotka <bwplotka@gmail.com>
This is technically BREAKING CHANGE, but it was like this from the beginning: I just notice that we rely in
Prometheus on remote read being sorted. This is because we use selected data from remote reads in MergeSeriesSet
which rely on sorting.
I found during work on https://github.com/prometheus/prometheus/pull/5882 that
we do so many repetitions because of this, for not good reason. I think
I found a good balance between convenience and readability with just one method.
Smaller the interface = better.
Also I don't know what TestSelectSorted was testing, but now it's testing sorting.
Signed-off-by: Bartlomiej Plotka <bwplotka@gmail.com>
* Fix bug with WAL watcher and Live Reader metrics usage.
Calling NewXMetrics when creating a Watcher or LiveReader results in a
registration error, which we're ignoring, and as a result other than the
first Watcher/Reader created, we had no metrics for either. So we would
only have metrics like Watcher Records Read for the first remote write
config in a users config file.
Signed-off-by: Callum Styan <callumstyan@gmail.com>
This fixes#6992, which was introduced by #6777. There was an
intermediate component which translated TSDB errors into storage errors,
but that component was deleted and this bug went unnoticed, until we
were watching at the Prombench results. Without this, scrape will fail
instead of dropping samples or using "Add" when the series have been
garbage collected.
Signed-off-by: Julien Pivotto <roidelapluie@inuits.eu>
With defer having less of a performance penalty, there is no reason
not to do those crucial operations via defer.
Context: With isolation in place, if we forget to Commit/Rollback, the
low watermark will get stuck forever.
The current code should not have any bugs, but moving to defer helps
to avoid future bugs.
This is also moving the `closeAppend` in the `Commit` implementation
itself to defer. If logging to the WAL fails, we would have missed the
`closeAppend`.
Signed-off-by: beorn7 <beorn@grafana.com>
This is technically BREAKING CHANGE, but it was like this from the beginning: I just notice that we rely in
Prometheus on remote read being sorted. This is because we use selected data from remote reads in MergeSeriesSet
which rely on sorting.
I found during work on https://github.com/prometheus/prometheus/pull/5882 that
we do so many repetitions because of this, for not good reason. I think
I found a good balance between convenience and readability with just one method.
Smaller the interface = better.
Also I don't know what TestSelectSorted was testing, but now it's testing sorting.
Signed-off-by: Bartlomiej Plotka <bwplotka@gmail.com>
I think the previous behavior is problematic as it will leave
`memSeries` around that still have `pendingCommit` set to `true`.
The only case where this can happen in this code path is a failure to
write to the WAL, in which case we are probably in trouble anyway. I
believe, however, we should still try to do the right thing and do the
full rollback. This will implicitly try to write to the WAL again, but
this time without samples, which may even succeed. (But we propagate
the previous error in any case.)
This also adds `a.head.putSeriesBuffer(a.sampleSeries)` to Rollback,
which was previously missing.
Signed-off-by: beorn7 <beorn@grafana.com>
This is taken from #6918. Since we probably won't merge #6918 before
the relase, we have to do this bit of it as it fixes an actual bug
(iso.closeAppend is not called if the append fails because of an error
logging to the WAL).
Signed-off-by: beorn7 <beorn@grafana.com>
* tsdb: don't allow ingesting empty labelsets
When we ingest an empty labelset in the head, further blocks can not be
compacted, with the error:
```
level=error ts=2020-02-27T21:26:58.379Z caller=db.go:659 component=tsdb
msg="compaction failed" err="persist head block: write compaction:
add series: out-of-order series added with label set \"{}\" / prev:
\"{}\""
```
We should therefore reject those invalid empty labelsets upfront.
This can be reproduced with the following:
```
cat << END > prometheus.yml
scrape_configs:
- job_name: 'prometheus'
scrape_interval: 1s
basic_auth:
username: test
password: test
metric_relabel_configs:
- regex: ".*"
action: labeldrop
static_configs:
- targets:
- 127.0.1.1:9090
END
./prometheus --storage.tsdb.min-block-duration=1m
```
And wait a few minutes.
Signed-off-by: Julien Pivotto <roidelapluie@inuits.eu>
Series() will fetch all the metadata for a series,
even if it's going to be filtered later due to time ranges.
For 1M series we save ~1.1s if you only needed some of the data, but take an
extra ~.2s if you did want everything.
benchmark old ns/op new ns/op delta
BenchmarkHeadSeries/1of1000000-4 1443715987 131553480 -90.89%
BenchmarkHeadSeries/10of1000000-4 1433394040 130730596 -90.88%
BenchmarkHeadSeries/100of1000000-4 1437444672 131360813 -90.86%
BenchmarkHeadSeries/1000of1000000-4 1438958659 132573137 -90.79%
BenchmarkHeadSeries/10000of1000000-4 1438061766 145742377 -89.87%
BenchmarkHeadSeries/100000of1000000-4 1455060948 281659416 -80.64%
BenchmarkHeadSeries/1000000of1000000-4 1633524504 1803550153 +10.41%
benchmark old allocs new allocs delta
BenchmarkHeadSeries/1of1000000-4 4000055 28 -100.00%
BenchmarkHeadSeries/10of1000000-4 4000073 87 -100.00%
BenchmarkHeadSeries/100of1000000-4 4000253 630 -99.98%
BenchmarkHeadSeries/1000of1000000-4 4002053 6036 -99.85%
BenchmarkHeadSeries/10000of1000000-4 4020053 60054 -98.51%
BenchmarkHeadSeries/100000of1000000-4 4200053 600074 -85.71%
BenchmarkHeadSeries/1000000of1000000-4 6000053 6000094 +0.00%
benchmark old bytes new bytes delta
BenchmarkHeadSeries/1of1000000-4 229192184 2488 -100.00%
BenchmarkHeadSeries/10of1000000-4 229193336 5568 -100.00%
BenchmarkHeadSeries/100of1000000-4 229204856 35536 -99.98%
BenchmarkHeadSeries/1000of1000000-4 229320056 345104 -99.85%
BenchmarkHeadSeries/10000of1000000-4 230472056 3894673 -98.31%
BenchmarkHeadSeries/100000of1000000-4 241992056 40511632 -83.26%
BenchmarkHeadSeries/1000000of1000000-4 357192056 402380440 +12.65%
Signed-off-by: Brian Brazil <brian.brazil@robustperception.io>
Guangwen Feng