Skip to content
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

Flush fix for SimpleLogProcessor #1308

Merged

Conversation

cijothomas
Copy link
Member

@cijothomas cijothomas commented Oct 19, 2023

SimpleLogProcessor's force_flush was incorrectly implemented as it did not do anything assuming there was nothing to flush. But the SimpleLogProcessor uses a queue with pub-sub mechanism to operate, so flush is not a no-op. This is handled correctly for SimpleSpanProcessor.

Instead of borrowing the approach from SimpleSpanProcessor, I have refactored the SimpleLogProcessor to use no queue and instead use a simple Mutex protected exporter/shutdown. I feel this is sufficient (and want to port this to SimpleSpanProcessor as well), but would like to get feedback on the approach - was there some scenario which prompted the queue/pub-sub for SimpleProcessors? Or this approach is sufficient?

There should not be any perf concerns as SimpleProcessors are used for learning/dev scenarios, and not for production. (Except when you are exporting to operating system native tracing like etw, user_events, but for them we have written RentrantProcessor separately.)

Note: I temporarily added some extra dependencies to test this out. If the direction is okay, I'll do a cleanup and modify the tests to not rely on tracing and friends.

@cijothomas cijothomas requested a review from a team October 19, 2023 14:40
@cijothomas
Copy link
Member Author

The approach used in #1305 was used here as well. Its proving very useful to catch issues quickly, so irrespective of the actual fix, expect a bunch of new tests based on InMemoryLogExporter !

@cijothomas
Copy link
Member Author

If this approach is good enough, that can remove the "crossbeam-channel = { version = "0.5", optional = true }" dependency from sdk completely as well?

@codecov
Copy link

codecov bot commented Oct 19, 2023

Codecov Report

Attention: 4 lines in your changes are missing coverage. Please review.

Files Coverage Δ
opentelemetry-sdk/src/logs/mod.rs 100.0% <100.0%> (ø)
opentelemetry-sdk/src/logs/log_processor.rs 10.3% <73.3%> (+10.3%) ⬆️

... and 9 files with indirect coverage changes

📢 Thoughts on this report? Let us know!

@lalitb
Copy link
Member

lalitb commented Oct 19, 2023

Thanks for picking this up, and the InMemoryExporter tests :) So there are two options here

Option 1: Keep the existing queue based approach, and fix the SimpleLogProcessor::force_flush to process the queue (as done for spans). Which mean tokio::log!() -> Logger::emit() -> LogProcessor::emit() will return immediately after writing to queue, and a separate background thread will process the queue. No locks, and lags in application thread. The queue is unbounded, which means it can continue to grow if the input throughput exceeds the processing capacity. But then, SimpleLogProcessor is not for production use.

Option 2: The approach in this PR to do a export from application thread, and so a mutex lock is required to ensure that Exporter::export() is not called concurrently from multiple application threads. Simple approach, and without external dependency on crossbeam-channel.

While most of the language implementation (probably all) go with Option 2, I don't have strong opinion against Option 1 if we address the flushing issue. Nothing wrong with the performance gain even for non-production usage :)

@cijothomas
Copy link
Member Author

Thanks for picking this up, and the InMemoryExporter tests :) So there are two options here

Option 1: Keep the existing queue based approach, and fix the SimpleLogProcessor::force_flush to process the queue (as done for spans). Which mean tokio::log!() -> Logger::emit() -> LogProcessor::emit() will return immediately after writing to queue, and a separate background thread will process the queue. No locks, and lags in application thread. The queue is unbounded, which means it can continue to grow if the input throughput exceeds the processing capacity. But then, SimpleLogProcessor is not for production use.

Option 2: The approach in this PR to do a export from application thread, and so a mutex lock is required to ensure that Exporter::export() is not called concurrently from multiple application threads. Simple approach, and without external dependency on crossbeam-channel.

While most of the language implementation (probably all) go with Option 2, I don't have strong opinion against Option 1 if we address the flushing issue. Nothing wrong with the performance gain even for non-production usage :)

I want to use Option 2 only, unless there is a strong reason not to.

  1. It keeps the code a lot simpler, and reduces dependencies.
  2. Is aligned with the spirit of SimpleProcessors original spec intent.
  3. If someone really wants the user thread to return quickly, but does not want to pay for the buffer either, then they can use BatchProcessor with capacity=1. (Not a valid scenario in my opinion!)

@lalitb
Copy link
Member

lalitb commented Oct 19, 2023

If someone really wants the user thread to return quickly, but does not want to pay for the buffer either, then they can use BatchProcessor with capacity=1. (Not a valid scenario in my opinion!)

that's a good point. Good to get feedback on the reason behind original implementation :)

@TommyCpp TommyCpp self-assigned this Oct 20, 2023
@TommyCpp
Copy link
Contributor

Pub-sub based approach first introduced in #502 to address the performance issue of simple span processors. One thing worth noting is currently users who don't use async runtime in Rust cannot use BatchSpanProcessor so there were some argument to keep the SimpleSpanProcessor relatively performant.

@@ -42,62 +42,32 @@ pub trait LogProcessor: Send + Sync + Debug {
/// emitted. If you find this limiting, consider the batch processor instead.
#[derive(Debug)]
pub struct SimpleLogProcessor {
sender: crossbeam_channel::Sender<Option<LogData>>,
shutdown: crossbeam_channel::Receiver<()>,
exporter: Mutex<Box<dyn LogExporter>>,
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why does this need to be wrapped in a Mutex? It seems a processor owns its exporter and the methods to call on it are &mut self, so who is thing shared with that requires the Mutex?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can I table this, and come back to this after this PR?

@cijothomas
Copy link
Member Author

Pub-sub based approach first introduced in #502 to address the performance issue of simple span processors. One thing worth noting is currently users who don't use async runtime in Rust cannot use BatchSpanProcessor so there were some argument to keep the SimpleSpanProcessor relatively performant.

Thanks for sharing the background!
#502 does not seem to addressing any end user needs. What I understand is that, that PR addresses the "must not block" part from the spec, just for spec compliance purposes.
Also, the spec does not require SimpleSpanProcessor to not-block/high-perf. It only says the API/SDK should not block, and leaves out the SpanProcessors.

"Some downstream SpanProcessors and subsequent SpanExporters called from this operation may be used for testing, proof-of-concept ideas, or debugging and may not be designed for production use themselves. They are not in the scope of this requirement and recommendation."

@jtescher Can you comment on any actual use cases on having a performant simple span processor? Or we did that just for spec compliance? If that is the case, would you be okay to revert that change, as spec does not require that from us.

@cijothomas
Copy link
Member Author

One thing worth noting is currently users who don't use async runtime in Rust cannot use BatchSpanProcessor so there were some argument to keep the SimpleSpanProcessor relatively performant.

This is something we need to address separately, anyways, do you agree? i.e We want to support users who do not use aync runtimes. (not just traces, but metrics too)

@hdost
Copy link
Contributor

hdost commented Oct 24, 2023

Thanks for sharing the background!
#502 does not seem to addressing any end user needs.

Not sure I would assume that.

@cijothomas
Copy link
Member Author

Thanks for sharing the background!
#502 does not seem to addressing any end user needs.

Not sure I would assume that.

Can you share some user needs around SimpleProcessor? The linked PR talks mostly about spec compliance...

@TommyCpp TommyCpp removed their assignment Oct 24, 2023
@cijothomas
Copy link
Member Author

@jtescher Could you comment on #1308 (comment), when you get a chance ? I am mostly looking to ensure we agree (or disagree) on the direction, which is:

  1. Keep SimpleLog/SpanProcessors "simple", and documented to end-users that this is not meant for Production scenarios.
    a. As a bonus, crossbeam-channel dependency can be removed from sdk.
  2. Users without an async runtime cannot use BatchProcessor (they cannot use Metrics Sdk either), but this is an issue we want to address irrespective of this PR!

In the absence of any real user needs for high-performing SimpleProcessor, we should strive to keep it simple! Happy to change my views, if there are actual use cases I am not aware of!

@jtescher
Copy link
Member

jtescher commented Nov 8, 2023

Yeah I don't think it needs to be as performant 👍 . The simple case should be for scripts / jobs / other sync activities which shouldn't produce a volume of spans where it is impactful.

@cijothomas
Copy link
Member Author

@TommyCpp @jtescher Please re-review. I have fixed handling the lock poisoning in latest commits.

The question from @shaun-cox about the very need of Mutex - Let me get back to it after this PR.

This PR now makes it easy to write semi-integration tests for Logging using in-memory-exporter, and I want to add more of that.

@cijothomas
Copy link
Member Author

@hdost Can you re-review as well as you have previously reviewed?

@TommyCpp
Copy link
Contributor

@hdost Can you re-review as well as you have previously reviewed?

cc @hdost for final review

@hdost hdost merged commit cb2d127 into open-telemetry:main Nov 12, 2023
14 checks passed
@cijothomas cijothomas deleted the cijothomas/log-simpleprocessorfix branch November 13, 2023 19:47
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

6 participants