Module soak 

Proving nothing grows without bound.

This is the failure that only appears in production, because it is the only one that needs hours to become visible. A stack that leaks a task per call is indistinguishable from a correct one for the length of any test somebody runs impatiently.

Flat, not merely bounded. A leak that fills a pool is still a leak; the pool only hides it until it becomes a problem, and then hides the cause too. So the assertion is that the reading at the end matches the reading at the start, within a tolerance for the noise a runtime genuinely has — not that it stayed under some ceiling.

Measured after settling, and the settling period is longer than it looks. Ending a call is not instantaneous, and the slow part is not teardown — it is the protocol. RFC 3261 §17 keeps a completed server transaction alive for Timer J, 64·T1, thirty-two seconds, so it can absorb a retransmitted request. For that whole time there is a task per call that has ended, and it is doing exactly what the RFC requires.

So a settle shorter than the longest transaction timer reports the specification as a leak. The first version of sipx’s own soak used five seconds and duly failed with “tasks grew from 5 to 305” after 300 calls — a number that looks exactly like a one-task-per-call leak and was not one. SETTLE_PAST_TIMERS is the floor.

Structs

Reading

A reading of the things that must not grow.

Soak

What a soak run found.

Tolerance

How much drift is noise rather than a leak.

Constants

SETTLE_PAST_TIMERS

The shortest settling period that does not accuse the protocol of leaking.

Functions

alive_tasks

The number of tasks alive in the current runtime.

open_descriptors

How many file descriptors this process holds.

resident_kb

Resident memory in kilobytes.

sample

Take a reading of the process now.

soak

Run load, then wait for things to settle, and report what grew.