paultanay commented on PR #1047: URL: https://github.com/apache/skywalking-banyandb/pull/1047#issuecomment-4224028552
> > Observed impact in this environment: > > > > 1. Write throughput improved vs RF=1 at RF=2 (+28.7%) and RF=3 (+27.6%). > > It does not make sense. More RF will reduce throughput theoretically. Could you double-check it? > > > 2. Query latency stayed close at RF=2 (p95 -2.4%, p99 -0.6% vs RF=1), but increased at RF=3 (p95 +13.0%, p99 +18.9%). > > It sounds reasonable; more RF will affect query performance because more data needs to be analyzed. > > > 3. Resource usage: > > > > * Write phase CPU increased with higher RF, mainly on liaison and data roles. > > * Read phase CPU was relatively stable. > > More RF should use more CPU and memory to handle data. Could you increase the query load? > > Observed impact in this environment: > > > > 1. Write throughput improved vs RF=1 at RF=2 (+28.7%) and RF=3 (+27.6%). > > It does not make sense. More RF will reduce throughput theoretically. Could you double-check it? > > > 2. Query latency stayed close at RF=2 (p95 -2.4%, p99 -0.6% vs RF=1), but increased at RF=3 (p95 +13.0%, p99 +18.9%). > > It sounds reasonable; more RF will affect query performance because more data needs to be analyzed. > > > 3. Resource usage: > > > > * Write phase CPU increased with higher RF, mainly on liaison and data roles. > > * Read phase CPU was relatively stable. > > More RF should use more CPU and memory to handle data. Could you increase the query load? You are right. My previous conclusion was based on older/lower-query runs. I reran with higher query load (query_workers=8, query_iterations=4000), 3 fresh rounds per RF, and aggregated the results. Mean vs RF=1: write throughput RF=2 -9.7%, RF=3 -16.0%; query p95 RF=2 +9.2%, RF=3 +4.7%; query p99 RF=2 +5.6%, RF=3 ~0.0%. Read CPU and RSS are slightly higher with higher RF. I also found setup instability during reruns (image-tag inconsistency) and addressed it locally; I will push this together with the remaining requested review fixes. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
