[config change] use MessageCommitMode when executing future head block messages #2705
Conversation
cla-bot
bot
added
the
s
|
I created a draft PR that exposes Stylus long term cache metrics that can be helpful when implementing tests in this PR. I didn't implement tests in my PR since it requires that long term caching is working properly, which is not true in the master branch 😬 In case you want to use what I developed you can get the changes from my branch into your branch, and then continue and implement the tests. |
…stylus-lt-cache-test
|
|
||
| // See if the item is in the long term cache | ||
| if let Some(item) = cache.long_term.get(&key) { | ||
| return Some(item.data()); | ||
| } | ||
|
|
||
| // See if the item is in the LRU cache, promoting if so | ||
| if let Some(item) = cache.lru.get(&key) { | ||
| let data = item.data(); | ||
| if let Some(item) = cache.lru.peek(&key).cloned() { |
There was a problem hiding this comment.
This codepath clones data twice: once here in the "get" and the other when returning item.data().
Cloning the entry_size_estimate_bytes is o.k., but we don't want to clone module and engine unnecessarily.
This is where rust gets you :)
There are probably some solutions that would avoid cloning result of the peek, but I think simplest would probably be if you can avoid cloning in item.data() because item itself is discarded right after.
There was a problem hiding this comment.
Tsahi pointed out that there's one more unnecessary clone, so that's not fixed yet, working on it :)
There was a problem hiding this comment.
passed item.module and item.engine without cloning to the returned Option, let me know if that checks out :)
magicxyyz
changed the title
Co-authored-by: Diego Ximenes Mendes <[email protected]>
arbitrator/stylus/src/cache.rs
Outdated
| } | ||
| cache.long_term_counters.misses += 1; |
There was a problem hiding this comment.
I think this should only be increased if long_term_tag is 1. Because that would mean "this should be in long term cache".
There was a problem hiding this comment.
good point, it will filter out API calls noise and we should be able to observe how many misses are there when a node starts up. Changed :)
| pub extern "C" fn stylus_get_lru_cache_metrics() -> LruCacheMetrics { | ||
| InitCache::get_lru_metrics() | ||
| pub extern "C" fn stylus_get_cache_metrics() -> CacheMetrics { | ||
| InitCache::get_metrics() |
There was a problem hiding this comment.
@diegoximenes
this is a bug in the previous PR as well.
You're allocating memory here in rust and returning the pointer to go.
Go discards it because it's a garbage collected language and the memory is never released.
Solution is is to allocate CacheMetrics in go, pass a pointer and let rust update the data in the struct. That way rust doesn't allocate anything new and no memory is lost.
Fixes NIT-2812
Pulls: OffchainLabs/go-ethereum#362
Includes: #2712
This PR:
Fixes use of MessageRunMode values, so as MessageCommitMode is used when, and only when, the message is part of a soon-to-be head block. Previously, newly sequenced / synced messages were executed in MessageReplayMode - newly activated / set-cached stylus programs were not cached in long term cache (only in LRU).
Improves repopulating of long term cache after node restart - if program is onchain marked as cached, if its wasm is found in LRU then it is also added to long term cache. That can happen e.g. when a ephemeral call to cached program precedes its onchain execution.
Adds tests for stylus long term cache + for repopulating long term cache from LRU cache.
Adds metrics for Stylus long term cache (merged from Diego's draft: Stylus cache improvements #2712)
Adds config to disable collection of Stylus metrics from Go side (also from the Diego's draft)