Nesting multiple decorators breaks waiting times about backoff

[bgreen-litl]

Woah, I just looked at my own at example more carefully and it's weirder than that. The KeyError wait generator is getting reset each time. I think combining decorators doesn't work as neatly as was supposed. :/

[bgreen-litl]

You are seeing the effect of the default jitter behavior backoff.full_jitter. See documentation here. You can read more here. If you don't want that behavior you can use a different jitter strategy, or pass jitter=None for no jitter at all.

[martinvol]

Thanks a lot for your answer! Should I submit a PR to update this section of the docs?

I find this behavior is not intuitive (at least for me).

[bgreen-litl]

So I think your original report was correct, and I was confused because full_jitter was on. Full jitter can cause the value to go down from one retry to the next, but the behavior your were citing is real independent of jitter.

Here's an example with jitter disabled:

In [7]: @backoff.on_exception(backoff.expo, ValueError, jitter=None)
...: @backoff.on_exception(backoff.expo, KeyError, jitter=None)
...: def foo(calls=[ValueError(), KeyError(), ValueError(), KeyError(), Value
...: Error(), KeyError(), ValueError()]):
...: raise calls.pop()
...:

In [8]: foo()
Backing off foo(...) for 1.0s (ValueError)
Backing off foo(...) for 1.0s (KeyError)
Backing off foo(...) for 2.0s (ValueError)
Backing off foo(...) for 1.0s (KeyError)
Backing off foo(...) for 4.0s (ValueError)
Backing off foo(...) for 1.0s (KeyError)
Backing off foo(...) for 8.0s (ValueError)

You can see that each exception type gets its own wait sequence. This is because each decorator defines and instantiates its own wait generator. I think this is probably unavoidable, and probably(?) desirable. Consider if you wanted to do exponential backoff for one type of exception and constant backoff for another.

I think you're right that at the least it could be documented better. At of curiosity, do you have a real world case where you wanted to define different behavior for different exceptions?

[martinvol]

I think you're right that at the least it could be documented better. At of curiosity, do you have a real world case where you wanted to define different behavior for different exceptions?

Yes, just hitting an API and if I get an exception related with bad formatting I don't want to retry, but if I hit any other exception I want to back off exponentially.