mongodb-js / devtools-shared / 28511749931

79%

main: 79%

Pull #809

github

fix(devtools-proxy-support): serialize SSH reinit and cap consecutive failures HELP-96049

When the SSH tunnel breaks during hibernate, concurrent _connect() callers all
hit initialize(reinitializeClient=true) simultaneously. Because that path
bypasses the connectingPromise deduplication guard, every caller raced through
the reinit — each calling sshClient.end() + createSshClient() + connect() —
producing N new server-side SSH sessions per MongoDB driver retry cycle. With
the driver retrying every ~10 s indefinitely, sessions accumulated without
bound (300+ observed in the field, exhausting the bastion's RAM).

Two changes:

1. Serialize concurrent reinit attempts via reinitializingPromise. The first
   caller to reach initialize(true) owns the reinit; subsequent concurrent
   callers await that same promise and then fall through to initialize(false)
   to use the freshly-connected client. Only one new SSH session is opened per
   reinit cycle regardless of how many _connect() callers are in flight.

2. Cap consecutive reinit failures. A new consecutiveReinitFailures counter
   increments each time _doReinit() fails and resets to zero on success. After
   MAX_CONSECUTIVE_REINIT_FAILURES (3) consecutive failures the agent sets
   closed=true, causing all subsequent initialize() calls to throw
   'Disconnected.' immediately. This surfaces a clean error to the MongoDB
   driver rather than looping indefinitely and leaking sessions.

1921 of 2692 branches covered (71.36%)

Branch coverage included in aggregate %.

4127 of 4956 relevant lines covered (83.27%)

639.06 hits per line