tighten the minimal streams per 100ms for staked node #738
Conversation
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## v1.17 #738 +/- ##
=======================================
Coverage 81.6% 81.6%
=======================================
Files 806 806
Lines 219284 219303 +19
=======================================
+ Hits 179014 179072 +58
+ Misses 40270 40231 -39 |
| Percentage::from(MAX_UNSTAKED_STREAMS_PERCENT) | ||
| .apply_to(MAX_STREAMS_PER_100MS) | ||
| .saturating_div(MAX_UNSTAKED_CONNECTIONS as u64) | ||
| }; | ||
|
|
||
| let min_staked_streams_per_100ms = if max_unstaked_connections == 0 { |
There was a problem hiding this comment.
max_unstaked_connections == 0 please tell me that this isn't a sentinel for tpufwd port
There was a problem hiding this comment.
It is. But what it really means is also on that port we allow 0 unstaked connections and only staked ones. So literally it is correct. It does not matter if it is used for forward or not.
There was a problem hiding this comment.
i know what it means. that's not the problem. the problem is that it is extremely brittle logic that can be broken from other areas of the code base with no knowledge of its existence. we cannot have code like this
Problem
The small staked's minimal streams per second is too generous. The one barely be entitled to 10 PPS can get 80 making it effective in QOS.
Summary of Changes
This change achieve the following results which matches master's behavior
1, On the tpu forward port where unstaked connections is 0, set the staked node's minimal PPS to 10 instead of 80
2. On the tpu port, set the staked nodes's minimal streams per 100MS to the the one of unstaked connection's + 1 in 100MS. This encourage more staked vs unstaked.
Fixes #