krzysztof/landing_page
1
0
Fork 0
mirror of https://github.com/qdrant/landing_page.git synced 2026-10-10 21:38:30 +02:00
Code Issues Packages Projects Releases Wiki Activity

Read-write contention troubleshooting guide (#2421)

Browse Source 
* New r/w contention troubleshooting guide

* Add snippet to low latency search guide

* Add FAQ entry

* Add 'Read More' section to guide
This commit is contained in:
Abdon Pijpelink
2026-06-23 16:04:13 +02:00
committed by GitHub
parent 8dce906b74
commit 7590063d83
64 changed files with 926 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
qdrant-landing/content/documentation/search/low-latency-search.md
+7 -1
View File
@@ -36,7 +36,13 @@ By default, a search operation queries a single replica of each shard in a colle
To reduce tail latency for read operations, Qdrant supports delayed fan-outs. With delayed fan-outs, if the initial request to a replica exceeds a specified latency threshold, an additional read request is sent to another replica. Qdrant will then use the first available response.
You can enable delayed fan-outs per collection by [setting](/documentation/manage-data/collections/#update-collection-parameters) the `read_fan_out_delay_ms` parameter to the number of milliseconds to wait before attempting to read from another replica. To disable delayed fan-outs after enabling, set this parameter to `0` (default).
You can enable delayed fan-outs per collection by [setting](/documentation/manage-data/collections/#update-collection-parameters) the `read_fan_out_delay_ms` parameter to the number of milliseconds to wait before attempting to read from another replica.
{{< code-snippet path="/documentation/headless/snippets/update-collection/read-fan-out-delay-ms/" >}}
Replace `100` with your collection's measured p95 read latency.
To disable delayed fan-outs after enabling, set this parameter to `0` (default).
<aside role="alert">Do not set the latency threshold to a very low value (for example, the 5th percentile of the read latency). This would cause almost every request to trigger additional read requests, significantly increasing the load on the cluster without much benefit. A good starting point is to set it to the 95th percentile of the read latency. This limits the additional load to approximately 5% while substantially shortening the latency tail.</aside>