pubsub: expose a max_duration_per_lease_extension setting for subscribers

[hongalex]

For languages that support lease management, we have bound the percentile distribution to 10s - 10m. It would be useful if the upper bound were configurable.

A user might process messages very slowly (8m to ack)
If the application fails 5s after receiving a message (whose ack deadline was set to 8m, because that's the deadline the percentile distribution gave), the user has to wait 7m55s for the message to be redelivered
Users should be able to set max_duration_per_lease_extension = 20s, which caps an individual modAck RPC to 20s. This would have the downside of causing more modack RPCs to be sent, but the upside that the redelivery time would only be 15s
This setting differs from max_lease_duration, which defines the total amount of time we'll hold a message.

In addition, max_duration_per_lease_extension should be disabled if <=0, and should default to 0.

[IlyaFaer]

@pradn, that was me, who did this for Go, and, if you don't mind, I'll do this for Python as well

[pradn]

@IlyaFaer Sounds good, Ilya. Reach out if you have any questions.

[IlyaFaer]

@pradn, yes, I'm taking a look at ack_deadline() property and I don't see where _last_histogram_size value is changed. This attribute represents the size of the histogram on the last ack_deadline call, right?

I found only two places, where it's used:

python-pubsub/google/cloud/pubsub_v1/subscriber/_protocol/streaming_pull_manager.py

Line 111 in 380cb38

self._last_histogram_size = 0

python-pubsub/google/cloud/pubsub_v1/subscriber/_protocol/streaming_pull_manager.py

Line 194 in 380cb38

target = min([self._last_histogram_size * 2, self._last_histogram_size + 100])

Am I missing something, or target in here will always be equal to 100? Looks like something that needs to be fixed (I think, I'll push it by the way)

[hongalex]

Hey @IlyaFaer, we finalized the naming for all languages and updated the original issue. Please update when you get the chance.

[IlyaFaer]

@hongalex, @pradn, okay, it's done