Improve frameLag calculation to handle spikes and faster recovery #559
Conversation
|
(I'll look into fixing format errors) |
for formatting, you can just use |
|
I'm getting some weird errors on my end for compiling. I think it has to do with my Java version, I have Java 8, maybe I should switch to a higher JDK since it uses openjdk21. Running |
|
@lenguyenthanh Apologies for the wait, I just needed to update my Java. |
No need to apology and thanks for your contribution! So nice that it passes the CI. I'll let other to check if it worth merging. |
src/main/scala/Lag.scala
Outdated
|
|
||
| export trustedStats.getIfPresent as sessionLag | ||
|
|
||
| private val clientReports = groupedWithin[(User.Id, Int)](256, 947.millis): lags => | ||
| private val clientReports = groupedWithin[(User.Id, Int)](256, 947.millis) { lags => |
There was a problem hiding this comment.
please don't revert this syntax
src/main/scala/Lag.scala
Outdated
|
|
||
| export clientReports.apply as recordClientLag | ||
|
|
||
| def recordTrustedLag(millis: Int, userId: Option[User.Id], domain: Domain) = | ||
| Monitor.lag.roundFrameLag(millis, domain) | ||
| userId.foreach: uid => | ||
| userId.foreach { uid => |
| (prev * (1 - trustedRefreshFactor) + millis * trustedRefreshFactor).toInt | ||
| sessionLag(uid).fold(millis) { prev => | ||
| val weight = | ||
| if millis < prev then trustedNormalRefreshFactor else trustedSpikeRefreshFactor |
There was a problem hiding this comment.
This isn't about spikes, it's just about the current lag being any worse than the previous one.
So we would always be twice as slow recording higher lag than lower lag.
I expect that as a result our lag estimation would consistently be under-evaluated?
Implemented ideas as mentioned in #395 where massive spikes can cause the weighted average to skyrocket and take a long time to recover. The changes introduce separate refresh factors for spike and normal values, cap the maximum recorded value, and adjust the calculation to allow faster recovery when recording lower values.