distributor: Expermental support for producing ingest-storage records in format V2 #12060
Conversation
pkg/mimirpb/symbols_test.go
Outdated
| } | ||
|
|
||
| func BenchmarkSymbolizer(b *testing.B) { | ||
| b.Run("prom symbolizer: 10k labels unique values", func(b *testing.B) { |
There was a problem hiding this comment.
Individual sends from distributors to ingesters typically have 7 series, each with say 20 labels, so 280 strings total. That's the sort of ball-park we should be benchmarking.
There was a problem hiding this comment.
Agree on few series per per-partition write request. Let's say max_samples_per_send is 2000 and requests are full. Then we have per-partition write requests within this range:
- A tenant with shard size = 1: 2000 symbols in the per-partition write request
- A tenant with shard size = all ingesters: with 100 partitions that's 200 samples per per-partition write request, with 300 partitions (which is pretty large scale) that's 7, the number Bryan mentioned
There was a problem hiding this comment.
Gotcha, thank you for the context.
I want to move the benchmark up a level, using one benchmark to cover the whole flow for writes including both marshalling and unmarshalling. When I do that, I'll also make the request look more realistic.
There was a problem hiding this comment.
I've updated this benchmark to be more realistic (a case for both bryan's and marco's scenarios). It's also not coupled to the implementation so much anymore, thus easier to use with benchstat and iterate on.
alexweav
force-pushed
the
alexweav/rwv2-symbolizer
branch
from
e4581bb to
66c02a9
Compare
What this PR does
This PR adds experimental support for
ingest-storage.kafka.producer-record-version=2. Consumer side support exists already, this allows toggling producers to write the new format.This implements translation of RW1.0 to the new RW2.0-esque warpstream wire format.
We are not considering the format 100% stabilized, still. We continue to reserve the right to make backward-incompatible changes to the format, particularly the common symbols space.
In order to keep the PR size reasonable (it's already big), this PR omits a few key optimizations from main...alexweav/experiment-rw-v2 and focuses more on the plain implementation of the format. Namely, we don't yet pool the RW2.0 WriteRequest fields yet, similar to
PreallocTimeseries, nor do we use yoloStrings when marshalling. These will be saved for a future PR.The symbols table is an exception - it has been optimized to be significantly faster than the Prometheus implementation, with near-zero allocations.
Which issue(s) this PR fixes or relates to
contrib github.com/grafana/mimir-squad/issues/2253
Checklist
CHANGELOG.mdupdated - the order of entries should be[CHANGE],[FEATURE],[ENHANCEMENT],[BUGFIX]. If changelog entry is not needed, please add thechangelog-not-neededlabel to the PR.about-versioning.mdupdated with experimental features.