KAKafka · English
Community homeHot discussions Latest posts Best of Knowledge base Resources About

Kafka · English · Open-sourced a rate-limiting middleware for kafka-en with token bucket and sliding window

261
KAr/kafka-en·posted by huang_ke·3 days agoExperience

Open-sourced a rate-limiting middleware for kafka-en with token bucket and sliding window

Some background first. Our setup is kafka-en plus three downstream services, seven figures of daily requests, peaking around nine in the evening.

The first thing was to collapse the variables. We were changing config and upgrading the version at the same time, and afterwards nobody could say which change caused what. We rolled back to moving one variable at a time, re-ran three times, and only then did the curve settle. Tedious, but not skippable.

# Load test: do not jump straight to the max concurrency.
# Ramp it up, otherwise you miss the knee.
for c in 50 100 200 400 800; do
  wrk -t8 -c$c -d60s --latency http://127.0.0.1:8080/api/feed
  sleep 20
done

We also fixed monitoring along the way: replaced average-based alerts with percentiles and split them per endpoint. False alerts dropped by about seventy percent and the on-call rotation visibly cheered up.

One last trap: in container environments remember to adjust the memory-related parameters in step. Otherwise the host limit and the process expectation disagree, and the symptom is intermittent, unreproducible failure.

What genuinely surprised me was the tail. The average looked great while P99 jumped by an order of magnitude past some threshold. The cause was not kafka-en itself but our upstream connection reuse — the load test traffic was too clean and hid the long-tail requests.

57 comments

57 comments

M
Rrase·2 days agoedited

Saved. I am reworking this area this week — this saves a lot of wrong turns.

519
Ttang_hao·5 hours ago

This is not a kafka-en problem, it is a usage problem. The docs say this API is not thread-safe and you must lock around it yourself.

394
Wwinter·yesterday

Thanks for sharing real numbers — far more useful than the articles that only cover concepts.

300
Sslow_query·1 hour ago

I just read the kafka-en source — the author actually explains the reasoning in a comment, roughly "so that it degrades into predictable behaviour in extreme cases".

389
Sswoole_lee·1 hour ago

We have run this in production for two years without hitting it. That said, we never reached this scale, so our experience is not really evidence here.

439
Wwinter·5 hours ago

I see point 3 differently. The trade-off depends on your read/write ratio: read-heavy with little writing means caching actually widens the inconsistency window.

296
Sswoole_lee·3 minutes ago

Has anyone run a controlled experiment? I did, reducing it to a single variable, and the difference was 4% — within noise. So I suspect the main cause is something else.

200
Cchen_dev·2 hours ago

This matches what we see in production. We only hit it past 3k QPS; the earlier load tests showed nothing — the test traffic was too clean, with no long-tail requests.

259
Hhuang_ke·2 days ago

One counter-example: below kafka-en 7.4 the semantics of that code are different, so do not copy it verbatim. We got burned in staging and rolled back once.

431
Ttang_hao·5 hours agoedited

A question: what changes in a container with a 512Mi memory limit? That is how we run it in production.

159
Zzhou_yi·5 hours ago

I see point 3 differently. The trade-off depends on your read/write ratio: read-heavy with little writing means caching actually widens the inconsistency window.

88
Cchen_dev·just now

This is not a kafka-en problem, it is a usage problem. The docs say this API is not thread-safe and you must lock around it yourself.

39
Hhuang_ke·2 days ago

Saved. I am reworking this area this week — this saves a lot of wrong turns.

24
Kkite·2 days ago

Thanks for sharing real numbers — far more useful than the articles that only cover concepts.

425
Rran_bo·28 minutes ago

Agreeing with the above. One addition: with this option enabled the GC count in your metrics doubles, so adjust the alert threshold at the same time or it will keep firing.

1
Cchen_dev·2 days ago

This is not a kafka-en problem, it is a usage problem. The docs say this API is not thread-safe and you must lock around it yourself.

76
Mmike_xu·12 minutes agoedited

Worth learning from this debugging approach. We went straight at the logs and took a much longer route.

185
Wwinter·2 days ago

A question: what changes in a container with a 512Mi memory limit? That is how we run it in production.

368
Lli_ming·yesterday

Agreeing with the above. One addition: with this option enabled the GC count in your metrics doubles, so adjust the alert threshold at the same time or it will keep firing.

144
Hhuang_ke·2 days ago

We have run this in production for two years without hitting it. That said, we never reached this scale, so our experience is not really evidence here.

496
Sswoole_lee·2 days ago

Can you give a minimal reproduction? I ran it locally for ten minutes and could not reproduce on macOS with the latest version.

41
Sslow_query·12 minutes ago

Sharing our numbers, 8 cores 16GB, same scenario:

| Concurrency | P50 | P99 |
|---|---|---|
| 200 | 12ms | 88ms |
| 500 | 31ms | 340ms |

P99 clearly collapses at 500 concurrency, which lines up with your knee point.

118
Zzhou_yi·3 minutes ago

A question: what changes in a container with a 512Mi memory limit? That is how we run it in production.

103
Ttang_hao·1 hour ago

Saved. I am reworking this area this week — this saves a lot of wrong turns.

149
Zzhou_yi·3 minutes ago

Sharing our numbers, 8 cores 16GB, same scenario:

| Concurrency | P50 | P99 |
|---|---|---|
| 200 | 12ms | 88ms |
| 500 | 31ms | 340ms |

P99 clearly collapses at 500 concurrency, which lines up with your knee point.

5
Llinlin·2 days ago

This matches what we see in production. We only hit it past 3k QPS; the earlier load tests showed nothing — the test traffic was too clean, with no long-tail requests.

58
Nnikic·2 days ago

Sharing our numbers, 8 cores 16GB, same scenario:

| Concurrency | P50 | P99 |
|---|---|---|
| 200 | 12ms | 88ms |
| 500 | 31ms | 340ms |

P99 clearly collapses at 500 concurrency, which lines up with your knee point.

92
Ddev_zhou·3 minutes ago

There is actually a simpler fix that needs no architecture change: move this check up to the gateway and the problem disappears. The cost is one extra lookup at the gateway.

86
Zzhu_zong·28 minutes ago

Thanks for sharing real numbers — far more useful than the articles that only cover concepts.

3
Zzhu_zong·yesterday

A question: what changes in a container with a 512Mi memory limit? That is how we run it in production.

22
Hhuang_ke·just now

One counter-example: below kafka-en 7.4 the semantics of that code are different, so do not copy it verbatim. We got burned in staging and rolled back once.

13
Cchen_dev·2 days ago

One counter-example: below kafka-en 7.4 the semantics of that code are different, so do not copy it verbatim. We got burned in staging and rolled back once.

84
Hhuang_keOP·yesterday

This is not a kafka-en problem, it is a usage problem. The docs say this API is not thread-safe and you must lock around it yourself.

16
Ttang_hao·2 days ago

Worth learning from this debugging approach. We went straight at the logs and took a much longer route.

68
Lli_ming·1 hour ago

I just read the kafka-en source — the author actually explains the reasoning in a comment, roughly "so that it degrades into predictable behaviour in extreme cases".

252
Oops_wangOP·2 days ago

This matches what we see in production. We only hit it past 3k QPS; the earlier load tests showed nothing — the test traffic was too clean, with no long-tail requests.

1
Oops_wang·2 days ago

Agreeing with the above. One addition: with this option enabled the GC count in your metrics doubles, so adjust the alert threshold at the same time or it will keep firing.

50
Rran_bo·2 days agoedited

Thanks for sharing real numbers — far more useful than the articles that only cover concepts.

199
Llinlin·yesterday

There is actually a simpler fix that needs no architecture change: move this check up to the gateway and the problem disappears. The cost is one extra lookup at the gateway.

35
Kkite·3 minutes ago

Worth learning from this debugging approach. We went straight at the logs and took a much longer route.

116
Rrase·2 days ago

I see point 3 differently. The trade-off depends on your read/write ratio: read-heavy with little writing means caching actually widens the inconsistency window.

18
Bbob_chen·2 days ago

Sharing our numbers, 8 cores 16GB, same scenario:

| Concurrency | P50 | P99 |
|---|---|---|
| 200 | 12ms | 88ms |
| 500 | 31ms | 340ms |

P99 clearly collapses at 500 concurrency, which lines up with your knee point.

15
Lli_ming·2 days ago

Has anyone run a controlled experiment? I did, reducing it to a single variable, and the difference was 4% — within noise. So I suspect the main cause is something else.

8
Kkernel_panic·28 minutes ago

Can you give a minimal reproduction? I ran it locally for ten minutes and could not reproduce on macOS with the latest version.

7
Rran_bo·2 days ago

I just read the kafka-en source — the author actually explains the reasoning in a comment, roughly "so that it degrades into predictable behaviour in extreme cases".

27
Zzhu_zong·2 hours ago

Can you give a minimal reproduction? I ran it locally for ten minutes and could not reproduce on macOS with the latest version.

173
Rrase·2 days agoedited

Saved. I am reworking this area this week — this saves a lot of wrong turns.

306
Zzhou_yi·2 days agoedited

Worth learning from this debugging approach. We went straight at the logs and took a much longer route.

131
Zzhu_zong·2 days agoLevel 6

This matches what we see in production. We only hit it past 3k QPS; the earlier load tests showed nothing — the test traffic was too clean, with no long-tail requests.

490
Mmike_xu·2 days ago

There is actually a simpler fix that needs no architecture change: move this check up to the gateway and the problem disappears. The cost is one extra lookup at the gateway.

1
Llinlin·5 hours ago

One counter-example: below kafka-en 7.4 the semantics of that code are different, so do not copy it verbatim. We got burned in staging and rolled back once.

2
Ttang_hao·28 minutes ago

We have run this in production for two years without hitting it. That said, we never reached this scale, so our experience is not really evidence here.

5
Ddev_zhou·2 hours ago

Can you give a minimal reproduction? I ran it locally for ten minutes and could not reproduce on macOS with the latest version.

122
Oops_wang·5 hours ago

Agreeing with the above. One addition: with this option enabled the GC count in your metrics doubles, so adjust the alert threshold at the same time or it will keep firing.

3
Aalice_dev·28 minutes ago

Has anyone run a controlled experiment? I did, reducing it to a single variable, and the difference was 4% — within noise. So I suspect the main cause is something else.

70
Kkernel_panicOP·5 hours agoedited

I see point 3 differently. The trade-off depends on your read/write ratio: read-heavy with little writing means caching actually widens the inconsistency window.

4
Bbob_chen·2 days ago

There is actually a simpler fix that needs no architecture change: move this check up to the gateway and the problem disappears. The cost is one extra lookup at the gateway.

1

This is the post detail page /en/c/kafka-en/post/p13. Posts and comments are generated deterministically from a seeded PRNG, so the same post always renders the same content and the link can be shared, reloaded and indexed. In production this page reads MySQL for the post, Redis for hot-post caching, and fetches the whole comment tree in a single query on the path column.

See the database schema →