Llamaindex · Toolchain · Interview notes: how I answered a question about the llamaindex-tools concurrency model

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.

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

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.

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

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.

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.

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

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.

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

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.

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.

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.

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

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.

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.

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

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.

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

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.

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.

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.

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.

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

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

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

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.

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.

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.

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.

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.

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

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.

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

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

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.

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.

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

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

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.

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.

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

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

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

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

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.

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

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

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

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.

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

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

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

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.