---
title: "Comment by Ahmet Özel on \"Recall vs Latency: Tuning CockroachDB Vector Indexes in Production\" | Hashnode"
description: "The drop-and-recreate sweep measures each partition configuration on a newly built index. I would add a second phase after sustained inserts and deletes, using"
url: "https://hashnode.com/posts/recall-vs-latency-tuning-cockroachdb-vector-indexes-in-production/6aece8989a7c5e5bd5e10792/comment/6aed93f8b623060bb5e924dc"
---

[Discussion](https://hashnode.com/posts/recall-vs-latency-tuning-cockroachdb-vector-indexes-in-production/6aece8989a7c5e5bd5e10792)

Comment [Recall vs Latency: Tuning CockroachDB Vector Indexes in Production](https://hashnode.com/posts/recall-vs-latency-tuning-cockroachdb-vector-indexes-in-production/6aece8989a7c5e5bd5e10792)

[Ahmet Özel](https://hashnode.com/@ahmetozel)

[Nov 6](https://hashnode.com/posts/recall-vs-latency-tuning-cockroachdb-vector-indexes-in-production/6aece8989a7c5e5bd5e10792/comment/6aed93f8b623060bb5e924dc)

The drop-and-recreate sweep measures each partition configuration on a newly built index. I would add a second phase after sustained inserts and deletes, using the same query set, because an index’s steady-state partition shape can differ from its clean-build shape.

Record the write workload and maintenance state, then compare recall and latency after the churn settles and while it is active. That gives the partition-size choice an operational test beyond the initial CSV curve. It would also help distinguish a setting that works for a mostly static reference corpus from one that remains useful for frequently updated application context.

[Sai Kiran Kshatriya](https://hashnode.com/@saikirankshatriya)

[4d ago](https://hashnode.com/posts/recall-vs-latency-tuning-cockroachdb-vector-indexes-in-production/6aece8989a7c5e5bd5e10792/comment/6aed93f8b623060bb5e924dc#reply-6af7e28ceb19d60a221dacdc)

Thank you Ahmet.

That is a fair gap. The drop-and-recreate sweep only measures each configuration on a freshly built index, which is the best case. After enough inserts, deletes, splits and merges, the partition shape comes from the write history, and that can change both recall and latency. I'll add a churn phase. For each configuration, I'll keep the index after the clean-build sweep, run a recorded insert/delete workload, and repeat the same queries while the writes are running and again after things settle. I'll also recompute the exact top-K at each point, since the correct answers change after churn.

I'd push back on a few parts, though.

The churn profile can matter more than the partition setting. A churn benchmark is only as useful as how closely its write pattern matches production. Uniform deletes, oldest-first retention, one hot tenant, and new vectors far from the existing clusters can each rank the same configurations differently. Pick the wrong profile and you get a more detailed answer that is just as unrepresentative as the clean-build curve, and it looks more trustworthy.

So I'd present Phase 2 as a check of how sensitive a configuration is to churn, not as the new source of truth application context that changes often, the numbers after churn should drive the choice. I'll update the post once I have results.

[Ahmet Özel](https://hashnode.com/@ahmetozel)

[1d ago](https://hashnode.com/posts/recall-vs-latency-tuning-cockroachdb-vector-indexes-in-production/6aece8989a7c5e5bd5e10792/comment/6aed93f8b623060bb5e924dc#reply-6afb773aba8669b3bad2823a)

That pushback is fair. I should have framed the churn phase as sensitivity to a recorded workload, with the production write distribution as the eventual sizing input.

I would keep several named profiles and publish the configuration ranking under each, including a hot-tenant case. Recomputing exact top-K at each checkpoint is the important correction, because both the candidate set and the nearest-neighbour ground truth move while writes are active.
