From 193a958c8d18e2d237d4500d264fada10d30e52f Mon Sep 17 00:00:00 2001
From: JSONbored <49853598+JSONbored@users.noreply.github.com>
Date: Sat, 11 Jul 2026 07:12:06 -0700
Subject: [PATCH] fix(selfhost): retry the container health-status check, not
 just /ready (#5089)

#5085 fixed the /ready HTTP probe to retry instead of single-attempt-
failing, then immediately surfaced this second, previously-masked
instance of the same bug on the very next deploy: the Docker
container health-status check right below it is also a single,
non-retrying check, and caught the container in Docker's own
transient "starting" state (its healthcheck re-probes /ready on an
independent interval/start_period schedule, so it can still be
"starting" even after /ready itself already answered 2xx).

Retries only through "starting" -- any other non-healthy/running
status (unhealthy, exited, restarting, ...) fails immediately rather
than burning the full retry budget on something that will not
resolve on its own. Verified against both paths: starting -> healthy
resolves via retry, and an immediate unhealthy status fails fast
without waiting.
---
 scripts/selfhost-post-update-check.sh | 18 +++++++++++++++++-
 1 file changed, 17 insertions(+), 1 deletion(-)

diff --git a/scripts/selfhost-post-update-check.sh b/scripts/selfhost-post-update-check.sh
index cd85d386c..6750f92fd 100755
--- a/scripts/selfhost-post-update-check.sh
+++ b/scripts/selfhost-post-update-check.sh
@@ -51,7 +51,23 @@ if [ "$ready" -ne 1 ]; then
   exit 1
 fi
 
-status="$(docker inspect --format '{{if .State.Health}}{{.State.Health.Status}}{{else}}{{.State.Status}}{{end}}' "$container_id" 2>/dev/null || true)"
+# Same race as the /ready probe above, one layer down: Docker's OWN healthcheck (docker-compose.yml)
+# reports "starting" until its FIRST probe completes, which can still be true here even though /ready
+# already answered 2xx (the two checks run on independent schedules) -- a normal boot must not fail on
+# a state that is expected to resolve on its own within seconds. Only "starting" is worth waiting
+# through; any other non-healthy/running status (unhealthy, exited, restarting, ...) is a real problem
+# and fails immediately rather than burning the full retry budget on something that will not resolve.
+status=""
+for _ in $(seq 1 "$READY_RETRIES"); do
+  status="$(docker inspect --format '{{if .State.Health}}{{.State.Health.Status}}{{else}}{{.State.Status}}{{end}}' "$container_id" 2>/dev/null || true)"
+  if [ "$status" = "healthy" ] || [ "$status" = "running" ]; then
+    break
+  fi
+  if [ "$status" != "starting" ]; then
+    break
+  fi
+  sleep "$READY_RETRY_DELAY_SECONDS"
+done
 echo "selfhost post-update check: $SERVICE container status=$status"
 if [ "$status" != "healthy" ] && [ "$status" != "running" ]; then
   echo "error: expected healthy or running, got $status" >&2