You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Currently, the replay package in Pebble works to capture and replay write batches and ingestions - all write operations to pebble. The purpose of this is to measure compaction performance, specifically how effectively Pebble can keep up with paced/unpaced write traffic - traffic that is actually CockroachDB-representative. This addressed a common issue with Pebble-specific ycsb benchmarks that didn't stress the LSM in the same way CRDB does.
However, the replay package is mostly only good for measuring changes to WAL/compaction write throughput. The read side of this equation is currently missing; we don't capture or replay read operations (eg. iterator creation and iteration on top of the db, batches, or snapshots). Capturing these operation(s) at the cockroach/pebble interface level, and measuring their latencies and/or ops/sec in addition to everything else already in replay would basically get us a Pebble-level, CockroachDB-representative benchmarking suite.
This is an umbrella issue that consists of many smaller pieces of work; as design progresses on this project, we will turn this into a meta issue and link subtasks here.
The text was updated successfully, but these errors were encountered:
Currently, the
replay
package in Pebble works to capture and replay write batches and ingestions - all write operations to pebble. The purpose of this is to measure compaction performance, specifically how effectively Pebble can keep up with paced/unpaced write traffic - traffic that is actually CockroachDB-representative. This addressed a common issue with Pebble-specific ycsb benchmarks that didn't stress the LSM in the same way CRDB does.However, the replay package is mostly only good for measuring changes to WAL/compaction write throughput. The read side of this equation is currently missing; we don't capture or replay read operations (eg. iterator creation and iteration on top of the db, batches, or snapshots). Capturing these operation(s) at the cockroach/pebble interface level, and measuring their latencies and/or ops/sec in addition to everything else already in
replay
would basically get us a Pebble-level, CockroachDB-representative benchmarking suite.This is an umbrella issue that consists of many smaller pieces of work; as design progresses on this project, we will turn this into a meta issue and link subtasks here.
The text was updated successfully, but these errors were encountered: