When we retry after a 429, we get another 429.
I suspect a round-down issue.
In our scenario:
we do requests, and eventually get a 429 Too Many Request with Retry-After: 19
we wait 19 seconds, then try again:
—> We get a 429 again, with Retry-After: 0
Now, can it be the actual values were 19.4 and 0.4 for example and they were rounded down for the Retry-After in the headers?
4 лайка
Probably, should be an easy fix. Another starter task @erlend_sh ?
2 лайка
dinomite
(Drew Stephens)
28.Ноябрь.2018 16:10:50
3
I’m not entirely sure where this is happening, but I suspect it has to do with use of Lua’s tonumber() in rate_limiter.rb —quoth Redis’ EVAL documentation :
Lua has a single numerical type, Lua numbers. There is no distinction between integers and floats. So we always convert Lua numbers into integer replies, removing the decimal part of the number if any.
4 лайка
andrei
(Andrei Prigorshnev)
11.Январь.2021 17:02:52
4
Я занимаюсь этим багом и завтра отправлю pull-request.
1 лайк
andrei
(Andrei Prigorshnev)
12.Январь.2021 16:52:18
5
Оказалось, что проблема не в округлении в Lua-скрипте. Lua-скрипт уже получает округлённые значения.
Более того, проблема вообще не в округлении. Могут использоваться неот округлённые секунды, и в любом случае можно получить ответ 429 с заголовком Retry-After: 0.
Вот исправление основной проблемы:
master ← AndrewPrigorshnev:fix/sliding-window-end-time-in-rate-limiter
merged 06:26PM - 12 Jan 21 UTC
If the sliding window size is N seconds, then a moment at the Nth second should … be considered as the moment outside of the sliding window.
Otherwise, if the sliding window is already full, at the Nth second, a new call wouldn't be allowed, but a time to wait before the next call would be equal to zero, which is confusing.
In other words, the end of the time range shouldn't be included in the sliding window. Let's say we start at the second 0, and the sliding window size is 10 seconds. In the current version of Rate Limiter, this sliding window will be considered as a time range `[0, 10]` (including the end of the range), which actually is 11 seconds in length.
After this fix, the time range will be considered as `[0, 10)` (excluding the end of the range), which is exactly 10 seconds in length.
А вот исправление дополнительной проблемы, которая иногда может вызывать ту же ошибку:
master ← AndrewPrigorshnev:fix/use-the-same-time-moment-in-redis-calls
merged 08:09PM - 12 Jan 21 UTC
3 лайка