Should add a new (binary?) histogram to indicate whether we decided to prefetch or not. I think 50% rand() is fine for this, unless there are other startup experiments we want to do at the same time.
I guess we want to avoid this spilling on some channels. Which channels would it be suitable for? aurora + nightly? only nightly?
Aurora/nightly sounds good to me.
Maybe we should try no prefetch + no preload?
(In reply to Brian R. Bondy [:bbondy] from comment #3) > Maybe we should try no prefetch + no preload? Worth a try, but that requires being able to randomly remove prefetch files. Seems like something best done in own bug.
Specifically I'd like to see if no prefetch + no preload is faster than no prefetch + preload. So both have no prefetch.
Created attachment 633340 [details] [diff] [review] Patch v1.
Comment on attachment 633340 [details] [diff] [review] Patch v1. r+ if you change STARTUP_USING_PRELOAD to STARTUP_USING_PRELOAD_TRIAL or something. It's ok to reuse a histogram for a slightly different meaning, not ok to change histogram bucket distribution, it really confuses our frontend.
Created attachment 633507 [details] [diff] [review] Patch v2 Made histogram name change.