From 248099db8c583f6e80fda24f45db74361b757128 Mon Sep 17 00:00:00 2001 From: Jiri Stransky Date: Wed, 14 Jun 2017 15:24:46 +0200 Subject: Fix race conditions between containers In many occasions we had log directory initialization containers without `detach: false`, which didn't guarantee that they'll finish before the container depending on them will start using the log directory. This is now fixed by moving the initialization container one global step earlier, so that we can keep the concurrency when creating the log dirs. (Using `detach: false` makes paunch handle just one container at a time, and as such it can have negative performance impact.) For services which have their container(s) starting in step_1, initialization cannot be moved to an earlier step, so the solution here was to just add `detach: false`. As a minor related change, cinder DB sync container now mounts the log directory from host to put cinder-manage.log into the expected location. Change-Id: I1340de4f68dd32c2412d9385cf3a8ca202b48556 --- docker/services/cinder-scheduler.yaml | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) (limited to 'docker/services/cinder-scheduler.yaml') diff --git a/docker/services/cinder-scheduler.yaml b/docker/services/cinder-scheduler.yaml index 9d94b578..8199c34b 100644 --- a/docker/services/cinder-scheduler.yaml +++ b/docker/services/cinder-scheduler.yaml @@ -81,9 +81,8 @@ outputs: owner: cinder:cinder recurse: true docker_config: - step_3: + step_2: cinder_scheduler_init_logs: - start_order: 0 image: &cinder_scheduler_image list_join: - '/' -- cgit 1.2.3-korg