Raising maxPoolSize vs tuning maxIdleTimeMS for intermittent MongoDB pool exhaustion

Answer to the Core Question

For a workload that is quiet for a while and then spikes, the quickest way to keep latency low is to prevent idle connections from being evicted (set maxIdleTimeMS to 0 or a very high value). Once the pool has a warm base of sockets, a burst will be served immediately, regardless of maxPoolSize or maxConnecting. Increasing maxPoolSize and maxConnecting only helps if the pool has already been reduced by eviction; it does not change the fact that a burst will have to wait for new connections to be created.

In practice, the decision is usually:

• Disable eviction (maxIdleTimeMS=0) when you can afford a few more sockets to stay open and you want zero warm‑up latency.
• Keep a conservative maxPoolSize and raise maxConnecting only if you have a strict upper bound on total sockets that the server can handle.

Why this matters for burst‑after‑idle traffic

When idle sockets are evicted, the pool size drops to the number of connections that were actively in use just before the idle period. A burst that arrives after that drop must wait for the pool to grow again. The growth rate is limited by maxConnecting and by the time it takes to establish a new TCP/TLS handshake, which is typically several milliseconds. During that window, requests are queued or timed‑out.

Is Eviction Timing Guaranteed?

The MongoDB driver specification only states that an idle connection is considered for eviction after maxIdleTimeMS milliseconds of inactivity. The exact moment the eviction occurs is implementation‑specific and can be delayed by background housekeeping or OS scheduling. Therefore you cannot rely on a strict deadline; you can only approximate it by observing the driver’s internal timers.

Metrics & Log Events that Distinguish Eviction vs. Queuing

Use the following indicators to confirm the root cause:

• Pool statistics (e.g., poolSize, available, evictions, queuedAcquisitions) – a drop in poolSize followed by a spike in evictions points to idle‑eviction.
• Driver logs – look for lines such as connection pool evicted idle connection versus connection acquisition timed out. The former indicates eviction, the latter indicates queuing.
• Timing correlation – if the burst occurs immediately after an eviction log, the bottleneck is eviction; if the burst starts while poolSize is still high, the bottleneck is likely maxConnecting or maxPoolSize.

Practical Steps for Your Environment

1. Enable debug logging for the driver. For Node.js, set debug=true in the connection string or use console.log callbacks. For Java, set org.mongodb.driver.level=DEBUG in log4j.properties.
2. Record pool stats before, during, and after a known idle period. In Java, call MongoClient.getPoolStats(); in Node.js, access client.topology.poolStats via the monitor event.
3. Simulate a burst after a long idle. Pause the workload for > maxIdleTimeMS, then resume. Observe the latency and the number of evicted connections.
4. If latency spikes, set maxIdleTimeMS=0 (or a very high value) and rerun the test. Verify that the pool size remains stable and that burst latency drops.
5. If you still see queuing, increase maxConnecting (e.g., to 2× maxPoolSize) and check that new connections are created faster.

Missing Diagnostic Detail Needed?

To refine the recommendation further, please provide:

• The current values of maxPoolSize and maxConnecting in your configuration.
• The typical size of a burst (requests per second) and the duration of idle periods.
• Whether your driver uses TLS and if the handshake cost is significant.

These details will help determine if a moderate increase in maxPoolSize could be sufficient or if eviction control is the only viable solution.