🐛 e2e: stop racing the deadline in the stream failure check

handleStreamTicker compares updateFailure to updateMaxFailure before
incrementing it, so with updateMaxFailure 2 and a 1s interval the bouncer
gives up on the third consecutive failed poll, roughly 3s after the
endpoint starts failing. The check slept 2s and then polled for a 200 for
up to 15s, leaving about a second of margin, and once that window closes it
never reopens: a runner under load turns this into a 15s wait followed by a
failure.

Assert the 200 immediately after the endpoint starts failing, which is
always inside the window, and keep polling for the 403 that follows.
This commit is contained in:
mhx
2026-08-03 13:56:04 +02:00
parent 4bac63ca83
commit 99673b1087
+4 -3
View File
@@ -34,10 +34,11 @@ body() {
echo "[$SCENARIO] previously CIDR-banned IP must pass again once deletion is polled"
wait_for_status "http://127.0.0.1:${WEB_PORT}/foo" 200 15 -H "X-Forwarded-For: 10.0.0.1"
echo "[$SCENARIO] making the stream endpoint fail -> bouncer must pass for one more cycle (updateMaxFailure: 2)"
echo "[$SCENARIO] making the stream endpoint fail -> bouncer must still pass (updateMaxFailure: 2)"
lapi_set_stream_fail
sleep 2 # update cache is every 1 seconds then waiting for minimum 1 cycle
wait_for_status "http://127.0.0.1:${WEB_PORT}/foo" 200 15 -H "X-Forwarded-For: 8.8.8.8"
# No sleep: the bouncer gives up on the third failed poll, so at a 1s interval the
# window where it still serves is ~3s wide and never reopens. Assert inside it.
assert_status "http://127.0.0.1:${WEB_PORT}/foo" 200 -H "X-Forwarded-For: 8.8.8.8"
echo "[$SCENARIO] bouncer must block everything (isStreamHealthy: false)"
wait_for_status "http://127.0.0.1:${WEB_PORT}/foo" 403 15 -H "X-Forwarded-For: 8.8.8.8"