lhotari commented on code in PR #23945: URL: https://github.com/apache/pulsar/pull/23945#discussion_r1973973887
########## pip/pip-406.md: ########## @@ -0,0 +1,152 @@ +# PIP-406: Introduce metrics related to dispatch_throttled_count + +# Background knowledge + +## Motivation + +Currently, users can monitor subscription backlogs using the `pulsar_subscription_back_log_no_delayed` metric. +However, if [dispatch throttling](https://pulsar.apache.org/docs/next/concepts-throttling/) is configured at the broker/topic/subscription level, +this metric may not accurately reflect whether the backlog is due to insufficient consumer capacity, as it could be caused by dispatch throttling. + +## Goals + +Introduce metrics to indicate the count of `messages/bytes throttled` for **broker/topic/subscription** level rate limit. +This allows users to write PromQL queries to identify subscriptions with high backlogs but low or no throttling, pinpointing backlogs caused by insufficient consumer capacity. + +## In Scope + +Broker Level: +- Introduce the metric `pulsar_broker_dispatch_throttled_msg_count` to represent the total count of messages throttled for a broker. +- Introduce the metric `pulsar_broker_dispatch_throttled_bytes_count` to represent the total count of bytes throttled for a broker. + +Topic Level: +- Introduce the metric `pulsar_dispatch_throttled_msg_count` to represent the total count of messages throttled for a topic. +- Introduce the metric `pulsar_dispatch_throttled_bytes_count` to represent the total count of bytes throttled for a topic. + +Subscription Level: +- Introduce the metric `pulsar_subscription_dispatch_throttled_msg_count` to represent the total count of messages throttled for a subscription. +- Introduce the metric `pulsar_subscription_dispatch_throttled_bytes_count` to represent the total count of bytes throttled for a subscription. + + +## Out of Scope +- These states are not persistent and will reset upon broker restart/ topic re-load / subscription reconnected. + +# High Level Design +1. Maintain `dispatchThrottleMsgCount` and `dispatchThrottleBytesCount` in `DispatchRateLimiter`. Increase these values in the `consumeDispatchQuota` method when the TokenBucket for messages or bytes is insufficient. +2. Output these fields when retrieving metrics. + + +# Detailed Design + +## Design & Implementation Details +1. Maintain `dispatchThrottleMsgCount` and `dispatchThrottleBytesCount` in `DispatchRateLimiter`: +```java + private final LongAdder dispatchThrottleMsgCount = new LongAdder(); + private final LongAdder dispatchThrottleBytesCount = new LongAdder(); +``` + +2. During each [consumeDispatchQuota](https://github.com/apache/pulsar/blob/c4cff0ab3dac169c0a1418ef2f63f61604f6278e/pulsar-broker/src/main/java/org/apache/pulsar/broker/service/persistent/DispatchRateLimiter.java#L97-L104), +if token bucket is insufficient, increase these fields accordingly. +```diff + public void consumeDispatchQuota(long numberOfMessages, long byteSize) { + AsyncTokenBucket localDispatchRateLimiterOnMessage = dispatchRateLimiterOnMessage; + if (numberOfMessages > 0 && localDispatchRateLimiterOnMessage != null) { +- localDispatchRateLimiterOnMessage.consumeTokens(numberOfMessages); ++ if (!localDispatchRateLimiterOnMessage.consumeTokensAndCheckIfContainsTokens(numberOfMessages)) { ++ dispatchThrottleMsgCount.increment(); ++ } + } + AsyncTokenBucket localDispatchRateLimiterOnByte = dispatchRateLimiterOnByte; + if (byteSize > 0 && localDispatchRateLimiterOnByte != null) { +- localDispatchRateLimiterOnByte.consumeTokens(byteSize); ++ if (!localDispatchRateLimiterOnByte.consumeTokensAndCheckIfContainsTokens(byteSize)) { ++ dispatchThrottleBytesCount.increment(); ++ } Review Comment: > If the record is here will get huge throttled count, Because once rate limiting occurs, there will be multiple repeated requests(asyncRead) in a short period. @shibd What I meant to say earlier is that if you handle the metric this way, it's completely bogus metric. The reason for this is that when the tokens are consumed, the messages have already been sent. The metric would be tracking sent messages and bytes and not throttled messages and bytes. I understand that when throttling occurs, the same messages might get throttled again. You would need to address that problem if you would like that this metric counts the same entry only once. Duplication in the throttling code was recently eliminated and there's now a single location in the code base where the dispatch rate limiter throttling happens, it's here: https://github.com/apache/pulsar/blob/efcf7c27b3f0fa5dc0252ce46f54b28a897e08bd/pulsar-broker/src/main/java/org/apache/pulsar/broker/service/AbstractBaseDispatcher.java#L417-L428 -- 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]
