LLLLM Community
Community homeHot discussions Latest posts Best of Knowledge base Resources About

LLM Community · After reading the llm core source I finally understand the trade-offs

73
LLr/llm·posted by ran_bo·3 hours agoReview

After reading the llm core source I finally understand the trade-offs

Short version: llm needs almost no tuning at small and medium scale — the point where it starts to hurt is much further out than most people assume. Full measurements below.

-- The query that broke: a full scan over 20M rows.
-- A composite index took P99 from 1.8s down to 42ms.
SELECT id, title, created_at
  FROM posts
 WHERE community_id = ?
   AND status = 1
 ORDER BY score DESC
 LIMIT 20;

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.

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.

24 comments

24 comments

M
Hhuang_ke·12 minutes ago

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

513
Bbob_chen·5 hours ago

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

509
Nnikic·2 days ago

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

407
Ddev_zhou·28 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.

335
Kkernel_panicOP·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.

341
Zzhou_yi·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.

351
Ddev_zhou·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.

227
Kkernel_panic·yesterday

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

65
Ddev_zhou·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.

199
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.

163
Kkite·2 days ago

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

132
Ddev_zhou·3 minutes ago

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

83
Ddev_zhou·2 days ago

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

50
Sslow_query·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.

29
Hhuang_ke·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.

8
Bbob_chen·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.

1
Nnikic·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.

517
Aalice_dev·yesterday

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.

260
Sswoole_lee·28 minutes ago

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

32
Ttang_haoOP·12 minutes ago

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

178
Wwinter·2 days ago

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

202
Cchen_dev·just nowLevel 6

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.

42
Wwinter·2 days agoedited

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

224
Sslow_query·1 hour ago

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

1

This is the post detail page /en/c/llm/post/p5. 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 →