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

Agent · Architecture · Cut our agent-design build from 4 minutes to 40 seconds with these 5 changes

491
AGr/agent-design·posted by ops_wang·1 hour agoPostmortem

Cut our agent-design build from 4 minutes to 40 seconds with these 5 changes

Most agent-design articles stop at "how to use it" and never cover "when not to use it". This is an attempt at the second half.

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

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 agent-design itself but our upstream connection reuse — the load test traffic was too clean and hid the long-tail requests.

On trade-offs, my view is this: if nobody on the team owns this area long-term, do not introduce a second mechanism. With two coexistence you first have to work out which one is even in play when things break, and that costs far more than the performance you saved.

100 comments

100 comments

M
Lli_ming·12 minutes 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".

478
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.

475
Nnikic·1 hour 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.

461
Rrase·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".

272
Sslow_query·1 hour ago

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

218
Lli_mingMod·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.

215
Zzhu_zong·2 days ago

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

1
Llinlin·2 days ago

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

45
Zzhou_yi·12 minutes 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.

428
Kkernel_panic·2 hours ago

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

56
Kkernel_panic·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.

309
Hhuang_ke·yesterday

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.

13
Mmike_xu·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.

49
WwinterOP·2 days ago

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

330
Sswoole_leeOP·2 days ago

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

305
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.

455
Ddev_zhou·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".

262
Aalice_dev·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".

346
Sslow_query·just now

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

35
Oops_wang·just now

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.

343
Zzhou_yi·2 days ago

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

2
Oops_wang·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.

309
WwinterOP·2 days ago

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

243
Sswoole_lee·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.

426
Llinlin·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.

486
Kkite·2 days 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.

305
Llinlin·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.

155
Ddev_zhouMod·yesterday

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

3
Hhuang_ke·12 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.

182
Lli_ming·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.

280
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.

273
Bbob_chen·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.

265
Nnikic·28 minutes ago

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

207
Zzhou_yiMod·yesterday

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

185
Hhuang_ke·yesterdayedited

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

513
Sswoole_lee·2 days ago

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

207
Kkernel_panicMod·2 hours 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.

195
Ttang_hao·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.

176
Cchen_dev·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.

165
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.

158
Llinlin·2 days agoedited

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

123
Kkite·2 days ago

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

132
Aalice_devOP·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.

49
Rran_bo·12 minutes ago

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

104
Bbob_chenMod·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.

93
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.

91
Sslow_query·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.

432
Nnikic·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".

357
Rran_bo·3 minutes 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.

1
Rran_bo·2 hours ago

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

1
Ttang_hao·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.

507
Kkite·2 days agoLevel 6

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

77
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.

255
Oops_wang·2 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.

59
Ddev_zhou·just nowedited

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

306
NnikicMod·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.

281
Hhuang_ke·3 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.

83
Ttang_hao·2 days ago

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

80
Sslow_query·3 minutes ago

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

49
Ddev_zhou·2 days ago

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

46
Nnikic·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.

46
Kkernel_panic·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".

45
Lli_ming·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.

38
Rrase·just now

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

37
Rrase·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.

36
Bbob_chenOP·just now

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

2
Oops_wangMod·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.

75
Ddev_zhou·just now

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
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.

365
Wwinter·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.

1
Kkernel_panic·2 days agoedited

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.

33
Sswoole_lee·28 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.

24
Sslow_query·1 hour 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
Sswoole_lee·12 minutes 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.

23
Zzhu_zong·just now

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.

124
Sslow_query·2 days agoedited

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.

47
Kkite·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.

1
Lli_ming·28 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.

425
Zzhu_zongOP·2 days 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.

500
Bbob_chenOP·2 days agoeditedLevel 6

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.

203
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.

451
Ddev_zhou·2 days agoeditedLevel 6

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.

7
Hhuang_ke·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.

128
WwinterOP·1 hour 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.

31
Llinlin·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.

67
Oops_wang·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.

8
Kkernel_panic·1 hour ago

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

32
LlinlinOP·2 days ago

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

428
Hhuang_ke·2 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.

1
Wwinter·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.

13
Wwinter·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.

25
Zzhu_zong·2 hours ago

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

5
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.

314
Ddev_zhou·12 minutes agoedited

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.

3
Kkite·2 days ago

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

2
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.

1
Llinlin·1 hour 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".

1
Zzhu_zong·2 days agoedited

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.

138
Bbob_chen·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
Zzhu_zong·2 days ago

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

1

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