From 7399fe2a07455d51fe983d80afe09d84d4846ca2 Mon Sep 17 00:00:00 2001 From: Matthew Meszaros Date: Fri, 13 Nov 2026 16:22:59 +0100 Subject: [PATCH] feat: describe in the AWS worker fleet guide the one case where lowering the fleet cap removes a worker, a scale-out racing the change, which skips the quiet period of the fleet scale-in --- docs/content/docs/development/aws-workers.mdx | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) diff --git a/docs/content/docs/development/aws-workers.mdx b/docs/content/docs/development/aws-workers.mdx index d94cdce34..826290c29 100644 --- a/docs/content/docs/development/aws-workers.mdx +++ b/docs/content/docs/development/aws-workers.mdx @@ -236,9 +236,10 @@ reaches the cap, and the backend logs `worker fleet is at its size cap` once an hour while it stays there. The new cap must be at least the running fleet and `fleet_min_size`, and at most -the ceiling, so lowering it never stops a running worker. If a scale-out lands -between that check and the change, the group scales in by that worker the way -any scale-in does: AWS first picks the Availability Zone with the most workers, +the ceiling. The one way lowering it stops a worker is a scale-out landing in the +moment between that check and the change: the group then removes one worker at +once, without the quiet period the fleet's own scale-in waits for, and picks it +the way any scale-in does: AWS first picks the Availability Zone with the most workers, then the newest worker in it, which can be one that holds mailboxes. It leaves through the drain hook and its mailboxes are placed on the rest of the fleet. A change asks for a fresh confirmation of the operator's identity and is recorded