gh-115168: Add pystats counter for invalidated executors #115169
Conversation
|
Does this count executors that are invalidated during interpreter shutdown, or have we turned off stats by then? |
Good call out. In the "normal" case where the user has turned on pystats directly, it would include the invalidations at shutdown, which are really teardowns, not invalidations. In the context of pyperformance and how we run benchmarks in our infrastructure, which turns pystats on/off around the benchmarking code, teardowns would be excluded from the number. (So the numbers I shared here should be correct). I suppose the better option would be to count these one layer up where we know why the executors are being invalidated. |
|
I've updated this to be more precise about when something is an invalidation vs. just a deallocation. |
Python/optimizer.c
Outdated
| @@ -1161,7 +1164,21 @@ _Py_Executors_InvalidateAll(PyInterpreterState *interp) | ||
| exec->vm_data.valid = false; | ||
| exec->vm_data.linked = false; | ||
| exec = next; | ||
| OPT_STAT_INC(executors_invalidated); |
There was a problem hiding this comment.
We still want stats for this, as it can be called when the builtins are modified, or other rare events.
Maybe add a flag parameter to distinguish between shutdown and actual invalidation?
There was a problem hiding this comment.
My approach was to move the counters to the callers, rather than a flag (but a flag is fine too if you prefer). In either case, it looks like I missed _PyInterpreterState_SetEvalFrameFunc: This counter should also be incremented there.
c5de632 to
db58a2f
Compare
|
@markshannon: This has had the merge conflicts fixed and is ready to merge. |
Counting when executors are invalidated seems like a useful statistic to have (see this discussion).