AGAgent · Architecture
Community homeHot discussions Latest posts Best of Knowledge base Resources About

Agent · Architecture · War story: agent-design timezone handling almost cost us a full day of billing

266
AGr/agent-design·posted by chen_dev·2 days agoPostmortem

War story: agent-design timezone handling almost cost us a full day of billing

It took me two weeks of on-and-off digging and plenty of wrong turns. Writing the process down as it happened so the next person spends less time.

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.

Documentation first44%
Source code first28%
Just ask someone17%
Run a demo and learn by error11%

452 votes total

63 comments

63 comments

M
Ddev_zhouOP·just now

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

506
Llinlin·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
Rrase·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.

98
Sslow_query·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.

432
Lli_ming·yesterday

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.

423
Cchen_devOP·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.

146
Ttang_hao·12 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.

417
Kkernel_panic·12 minutes ago

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

458
Sslow_query·2 days ago

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

95
Hhuang_ke·2 days ago

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

80
Kkernel_panic·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.

14
Oops_wang·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.

299
Hhuang_keOP·just now

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

268
Wwinter·2 days ago

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

261
Rran_bo·3 minutes agoedited

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

200
Wwinter·just nowedited

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

89
Ttang_hao·12 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.

63
Rran_bo·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.

27
Llinlin·2 days ago

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

178
Sswoole_lee·3 minutes ago

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

276
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.

491
Mmike_xu·just nowLevel 6

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.

411
Bbob_chen·12 minutes agoLevel 6

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.

128
Bbob_chenMod·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.

246
KkiteMod·2 days ago

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

104
Mmike_xu·5 hours agoLevel 6

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

336
Rran_bo·2 days agoLevel 6

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

9
Lli_mingMod·5 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.

2
Bbob_chen·2 days agoedited

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

414
Lli_ming·2 days agoedited

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.

267
Aalice_dev·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.

1
Ddev_zhou·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.

229
Sslow_query·2 days ago

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

88
Zzhu_zongOPMod·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.

263
Rrase·2 days agoedited

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.

153
Cchen_dev·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.

339
Kkernel_panic·1 hour 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.

133
Bbob_chen·12 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.

127
Zzhou_yi·28 minutes 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.

238
Kkernel_panic·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.

322
Zzhu_zong·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.

20
Bbob_chen·2 days ago

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

9
Cchen_devOP·2 days ago

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

46
Nnikic·2 days ago

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

495
Llinlin·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.

10
Ttang_hao·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.

20
Wwinter·2 days agoedited

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

101
Mmike_xuOP·2 days ago

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

1
Kkite·2 hours ago

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

71
Cchen_dev·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.

5
LlinlinOP·2 days ago

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

46
Ddev_zhouOP·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.

30
Aalice_dev·5 hours 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.

85
Nnikic·yesterday

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.

257
Ttang_hao·28 minutes ago

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

68
Oops_wang·5 hours agoedited

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

46
Wwinter·2 days ago

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

43
Llinlin·3 minutes 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.

39
Rrase·2 days ago

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

16
Llinlin·2 hours ago

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

5
Sslow_query·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.

1
Ttang_hao·2 days agoedited

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

1
Sslow_query·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.

185

This is the post detail page /en/c/agent-design/post/p4. 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 →