11 Commits (4c2feb32b55dc0431aff659607b67e4ac078325e)

Author	SHA1	Message	Date
Owen Diehl	05b8537e79	WAL Replay counter (#3974) ... * speccing out unordered head block * testware & unordered serialise * common utils for iter & sampleIter * more generic forEntries * more efficient unordedHeadChunk serialise (no internal re-casting) * roundtripping unordered head block, exit headchunk iteration early, add constant for current default chunk version * adds head block write benchmarks for ordered & unordered writes * fixes bench * removes unused initializer * gofmt * linting * introduces an entry counter into streams & wal records * sets & resets entrycount on checkpoint recovery and after wal replay finishes * pushing to streams checks counter * reject pre-applied entries test * removes commented test	4 years ago
Marco Pracucci	94041307b1	Upgrade Cortex (#3466) ... * Upgrade Cortex Signed-off-by: Marco Pracucci <marco@pracucci.com> * Updated vendor/modules.txt Signed-off-by: Marco Pracucci <marco@pracucci.com>	5 years ago
Owen Diehl	7613197593	prevents duplicate log lines from being replayed. closes #3378 (#3388)	5 years ago
Owen Diehl	224e68a50b	WAL backpressure (#3218) ... * adds WAL replay backpressure * backpressure fn doesnt lock during passed fn for concurrency * replayController * replayController supports a RemoveFlushedChunks fn * custom wal flusher * encoding test determinism * standardizes recoveredBytesTotal to replayControllers ownership * linting * default wal memory threshold to 4gb, adds docs * fix test after v3 chunk schema change * more lenient expectCheckpoint function * replayController protects against flush race condition * adds replay_memory_ceiling to wal jsonnet * replay ceiling help msg	5 years ago
Cyril Tovena	8716e24518	Allocate entries array with correct size while decoding WAL entries. (#3241) ... Small optimization I found while deployment the new version. We might want to have a pool of entries here. ``` ❯ benchcmp before.txt after.txt benchmark old ns/op new ns/op delta Benchmark_DecodeWAL-16 2350024 1591633 -32.27% benchmark old allocs new allocs delta Benchmark_DecodeWAL-16 20043 20005 -0.19% benchmark old bytes new bytes delta Benchmark_DecodeWAL-16 4946246 1763270 -64.35% ``` Signed-off-by: Cyril Tovena <cyril.tovena@gmail.com>	5 years ago
Cyril Tovena	07ece2bc2a	Improve checkpoint series iterator. (#3193) ... * Add basic benchmark. Signed-off-by: Cyril Tovena <cyril.tovena@gmail.com> * Improves memory usage of checkpointer series iterator. Signed-off-by: Cyril Tovena <cyril.tovena@gmail.com> * make lint. Signed-off-by: Cyril Tovena <cyril.tovena@gmail.com> * better size computation. Signed-off-by: Cyril Tovena <cyril.tovena@gmail.com> * Fixes test ordering flakyness. Signed-off-by: Cyril Tovena <cyril.tovena@gmail.com>	5 years ago
Cyril Tovena	0a81f70800	Improves checkpointerWriter memory usage (#3188) ... * Logs PushRequest data. This will allows to find information about received size and total entries per tenant. Example of a log from my dev testing: ``` level=debug ts=2021-01-15T11:16:21.735663076Z caller=http.go:67 org_id=3927 traceID=11c4774c6ec4bbf4 msg="push request parsed" path=/loki/api/v1/push content-type=application/x-protobuf body-size="11 kB" streams=5 entries=298 streamLabelsSize="1.9 kB" entriesSize="45 kB" totalSize="47 kB" ``` Of course this means we can use LogQL on this. Signed-off-by: Cyril Tovena <cyril.tovena@gmail.com> * Move metrics to avoid request traversal twice. Signed-off-by: Cyril Tovena <cyril.tovena@gmail.com> * Fixes tests. Signed-off-by: Cyril Tovena <cyril.tovena@gmail.com> * Set a base benchmark. Signed-off-by: Cyril Tovena <cyril.tovena@gmail.com> * Improves checkpointwriter memory usage. Signed-off-by: Cyril Tovena <cyril.tovena@gmail.com> * Kavi feedback. Signed-off-by: Cyril Tovena <cyril.tovena@gmail.com>	5 years ago
Cyril Tovena	2a90c4ceb5	Improve wal entries encoding. (#3153) ... Little trick to avoid memory allocation with regards to bytes slice and string. benchmp: ``` ❯ benchcmp before.txt after.txt benchmark old ns/op new ns/op delta Benchmark_EncodeEntries-16 1699362 1055627 -37.88% benchmark old allocs new allocs delta Benchmark_EncodeEntries-16 20025 25 -99.88% benchmark old bytes new bytes delta Benchmark_EncodeEntries-16 5625393 4665376 -17.07% ``` This originated from an investigation on CPU usage of ingester. Signed-off-by: Cyril Tovena <cyril.tovena@gmail.com>	5 years ago
Owen Diehl	4d9865acd4	Adds WAL support (experimental) (#2981) ... * marshalable chunks * wal record types custom serialization * proto types for wal checkpoints * byteswith output unaffected by buffer * wal & record pool ifcs * wal record can hold entries from multiple series * entry pool * ingester uses noopWal * removes duplicate argument passing in ingester code. adds ingester config validation & derives chunk encoding. * segment writing * [WIP] wal recovery from segments * replay uses sync.Maps & preserves WAL fingerprints * in memory wal recovery * wal segment recovery * ingester metrics struct * wal replay locks streamsMtx in instances, adds checkpoint codec * ingester metrics * checkpointer * WAL checkpoint writer * checkpointwriter can write multiple checkpoints * reorgs checkpointing * wires up checkpointwriter to wal * ingester SeriesIter impl * wires up ingesterRecoverer to consume checkpoints * generic recovery fn * generic recovery fn * recover from both wal types * cleans up old tmp checkpoints & allows aborting in flight checkpoints * wires up wal checkpointing * more granular wal logging * fixes off by 1 wal truncation & removes double logging * adds userID to wal records correctly * wire chunk encoding tests * more granular wal metrics * checkpoint encoding test * ignores debug bins * segment replay ignores out of orders * fixes bug between WAL reading []byte validity and proto unmarshalling refs * conf validations, removes comments * flush on shutdown config * POST /ingester/shutdown * renames flush on shutdown * wal & checkpoint use same segment size * writes entries to wal regardless of tailers * makes wal checkpoing duration default to 5m * recovery metrics * encodes headchunks separately for wal purposes * merge upstream * linting * addresses pr feedback uses entry pool in stream push/tailer removes unnecessary pool interaction checkpointbytes comment fillchunk helper, record resetting in tests via pool redundant comment defers wg done in recovery s/num/count/ checkpoint wal uses a logger encodeWithTypeHeader now creates its own []byte removes pool from decodeEntries wal stop can error * prevent shared access bug with tailers and entry pool * removes stream push entry pool optimization	5 years ago
Cyril Tovena	e6778e9827	Update to go 1.15 (#2877) ... * Upgrade to 1.15. Also upgrade linter and fixes some lint issues. Signed-off-by: Cyril Tovena <cyril.tovena@gmail.com> * Update CI. Signed-off-by: Cyril Tovena <cyril.tovena@gmail.com> * Update generated files. Signed-off-by: Cyril Tovena <cyril.tovena@gmail.com> * Moar lint 🤦 Signed-off-by: Cyril Tovena <cyril.tovena@gmail.com> * moar lint and tests fixes. Signed-off-by: Cyril Tovena <cyril.tovena@gmail.com> * Moar tests fixes. Signed-off-by: Cyril Tovena <cyril.tovena@gmail.com>	5 years ago
Owen Diehl	dc0bd01e35	WAL/marshalable chunks (#2764) ... * marshalable chunks * wal record types custom serialization * proto types for wal checkpoints * byteswith output unaffected by buffer	5 years ago