-
Notifications
You must be signed in to change notification settings - Fork 405
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
SpanAttributes modified to use Vec instead of OrderMap/EvictedHashMap #1293
SpanAttributes modified to use Vec instead of OrderMap/EvictedHashMap #1293
Conversation
I love the direction and think it's necessary to achieve the project's mission. |
I made a quick attempt to fix this issue for Spans on top of this branch, and the throughput reaches 16m/sec. Overall that'd be more than doubling the perf (6M/sec to 16M/sec). Of course I haven't thought through the full implications, so will wait for green light on this PR's direction before further exploring that... |
@cijothomas I think this makes sense, thanks for working on it. Retaining first n rather than last n seems fine so the perf win here is well worth it. We probably still want to keep track of the dropped counts where possible, could store vec and droppped count, or have a struct that stored both? |
I can bring it back. I left some comments in here about that feature overall : #1283 (comment). I'd pursue that separately, to avoid distracting this PR. Given there is +1s to the direction of this PR from other maintainer/approvers, I'll continue working on it. |
I think it's the right direction. We should be explicit in docs that attributions can duplicate so we don't need to have deduplication or sort logic for them. Thanks for working on it |
could also see where perf is with hashmap to keep de-duplication but change to last-write-wins, would be worth knowing the cost of each |
Yes thats what I tried first!. From my stress tests: 7.5 M/sec with regular HashMap (so sdk does dedup, but no eviction). 10 M/Sec with vec, where we removed de-duplication as well. (I do see a future where users will ask for de-duplication in the sdk level. We can do that as an opt-in SpanProcessor or even an Exporter feature, so as to no penalize every user.) |
Codecov ReportAttention:
... and 3 files with indirect coverage changes 📢 Thoughts on this report? Let us know!. |
@open-telemetry/rust-approvers Please review! |
Discussed with @cijothomas and we will follow up with the nits. Merging this for now |
Towards #1284 and #1283. Only focused on Span attributes (to be extended to Events, Links post this PR).
This PR modifes Span attributes to use Vec instead of OrderMap<Key, Value> in SpanBuilder( part of API), EvictedHashMap in Span (part of SDK). This is similar to what was done for Logs.
Performance Implications:
The performance gains are matching what is already demonstrated in #1121, but sharing the results from my run:
Stress test:
main branch:
Number of threads: 4
Throughput: 6,592,200 iterations/sec
Throughput: 6,504,400 iterations/sec
Throughput: 6,461,400 iterations/sec
This PR:
Number of threads: 4
Throughput: 10,274,200 iterations/sec
Throughput: 10,295,800 iterations/sec
Throughput: 10,329,000 iterations/sec
This is ~50% improvement in throughput!
Benchmarks:
start-end-span-4-attrs/always-sample
time: [586.89 ns 594.65 ns 603.28 ns]
change: [-38.279% -36.582% -34.871%] (p = 0.00 < 0.05)
Performance has improved.
This means we'll no longer be doing de-duplicating of span attributes. Opened an issue to track user feedback on this, which is also linked from changelog for easy visibility.
Additional Details:
The feature to enforce attributes limit is not done now. dropped_attribute_count will be 0 hence. It can be added back as a follow up, but I'll open a separate issue to discuss some issues in implementation of that, with some proposals.Have just done the bare minimum to show the changes and its perf implications. I expect this to be 1st of a series of PRs doing similar for Events, Links, and also bringing back the attribute limit enforcement, along with new unit tests coverage. I want to get an okay that this direction is agreed upon, before working further on this. (Note: There were few other potential ideas discussed in todays SIG call, which can further help perf, but I'll open separate issues to track them., and they can be additive changes after this). Also, there might be interest in doing de-dup on a opt-in basis, either in SDK or a custom Processor, or Exporter - this is another candidate for follow up!Thanks @shaun-cox and @lalitb for prior works, benchmarks and other consultation!