mirror of
https://github.com/KumoCorp/kumomta.git
synced 2026-08-23 00:00:14 +00:00
Wez Furlong
52 lines
2.4 KiB
Markdown
52 lines
2.4 KiB
Markdown
# strategy
|
|
|
|
{{since('2024.09.02-c5476b89')}}
|
|
|
|
Optional string to specify the scheduled queue strategy. There are two possible
|
|
values:
|
|
|
|
* `"TimerWheel"` - The timer wheel has `O(1)` insertion and `O(1)`
|
|
pop costs, making it good for very large scheduled queues, but that comes in
|
|
exchange for a flat periodic tick overhead. As the number of scheduled queues
|
|
increases and/or the `retry_interval` decreases, so does the aggregate overhead
|
|
of maintaining timerwheel queues.
|
|
* `"SkipList"` - A skiplist has `O(log n)` insertion and `O(1)` pop costs,
|
|
making it a good general purpose queue, but becomes more expensive to insert
|
|
into as the size of the queue increases. That higher insertion cost is in
|
|
exchange for the overall maintenance being cheaper, as the queue can go to
|
|
sleep until the time of the next due item. The ongoing and aggregate
|
|
maintenance is therefore cheaper than a `TimerWheel` but the worst-case
|
|
scenario where the destination is not accepting mail and causing the
|
|
scheduled queue to grow is logarithmically more expensive as the queue
|
|
grows.
|
|
* `"SingletonTimerWheel"` - Like `"TimerWheel"` above, but there is one shared
|
|
wheel for all queues that use the `"SingletonTimerWheel"` strategy. This
|
|
mitigates the per-queue aggregate maintenance overhead mentioned above, and
|
|
is ideal especially for deployments with very many (more than 100,000 or so)
|
|
scheduled queues. This option is the default in `dev` builds and will be
|
|
the default in the next stable release.
|
|
|
|
The default value depends on the version:
|
|
|
|
|Version|Default|
|
|
|-------|-------|
|
|
|2024.09.02-c5476b89 and earlier|`"TimerWheel"`|
|
|
|{{since('2024.11.08-d383b033', inline=True)}}|`"SingletonTimerWheel"`|
|
|
|
|
|
|
Which should you use? Whichever works best for your environment! Make sure that
|
|
you test normal healthy operation with a lot of queues as well as the worst
|
|
case scenario where those queues are full and egress is blocked.
|
|
|
|
If you have very short `retry_interval` set for the majority of your queues you
|
|
may wish to adopt `SkipList` for its lower idle CPU cost, or alternatively use
|
|
`TimerWheel` and find a `timerwheel_tick_interval` that works for your typical
|
|
number of queues.
|
|
|
|
!!! note
|
|
Changing the strategy for a given queue requires that the queue either be
|
|
aged out, or for kumod to be restarted, before it will take effect. This
|
|
restriction may be removed in a future release.
|
|
|
|
|