Repository navigation
Change backoff counter to have larger range and backoff more quickly. #141498
Description
The backoff counter, _Py_BackoffCounter, supports exponential backoff and a value up to 4095.
Backoff is exponential in powers 2.
There are two issues with this:
- It is a bit slow to backoff: specialization and JIT optimizer failures are rarely transient, so we are wasting effort with backoffs like 2, 4, 8,...
- It has a relatively low limit of 4095 forcing re-specialization/trace recording every 4095 iteration at most.
The fix is simple, increase the power from 2 to 4. This means we only need 3 bits for the backoff and we get a higher maximum of 8191.
We can also use a lookup table, to ensure that the number of iteration until the next specialization/jit is a prime number
| Backoff | Value | Iterations (value + 1) |
|---|---|---|
| 0 | 1 | 2 |
| 1 | 6 | 7 |
| 6 | 8190 | 8191 |
| 7 | UNREACHABLE_BACKOFF |
Activity
efimov-mikhail commentedon Nov 15, 2025
I'd like to work on this, if you're not against it, @markshannon.
efimov-mikhail commentedon Nov 15, 2025
I've created a PR with minimal fix: 12 + 4 bits -> 13 + 3 bits,
value = 2**backoff - 1->value = 2**(2*backoff+1) - 1.I can add using prime numbers to the same PR, if we want to.
CC @Fidget-Spinner
markshannon commentedon Nov 19, 2025
I did some work on it, but I'm happy for you to do it. It looks like you already are 🙂
gh-141498: Change backoff counter to use prime numbers instead of pow…
pythongh-141498: Change backoff counter to use prime numbers instead …
pythongh-141498: Change backoff counter to use prime numbers instead …