| 1 Jun 2025 |
 edef | 404s aren't cacheable, since presence isn't immutable | 17:43:23 |
 Arian | Aaaaaah | 17:43:33 |
| edef | and 404s tend to be unique requests | 17:43:35 |
| Arian | Yes. I see what you're getting at now | 17:43:45 |
| edef | every nix build is bottlenecked on cno 404s | 17:43:48 |
| edef | * every nix build is bottlenecked on c.n.o 404s | 17:43:52 |
| Arian | We don't cache them at all? Because we DO cache 404s locally | 17:44:52 |
| Arian | (found that out the hard way) | 17:44:57 |
| edef | we cache them at most briefly, but this isn't the point | 17:45:30 |
| edef | every derivation you build that isn't cached doesn't start building until it has been confirmed it 404s | 17:45:47 |
| edef | approximately 100% of all nix derivations aren't in the cache | 17:46:02 |
| edef | the caching is in the wrong direction | 17:46:29 |
| edef | caching absences buys you essentially nothing | 17:46:42 |
| edef | there are 1461501637330902918203684832716283019655932542976 (2^160) store paths, most of which are absent from the cache | 17:47:04 |
| edef | we only have 2^28 or so present store paths | 17:47:31 |
| edef | your 404 cache cannot fit 2^160 items | 17:47:41 |
| Arian | Bloom Filter? | 17:48:01 |
| Arian | Anyhow I digress. My idea definitely seems too simplistic | 17:48:12 |
| edef | yes, a bloom filter works | 17:48:18 |
 Vladimír Čunát | It's normal to cache even over larger spaces than 2^160. | 17:48:20 |
| Vladimír Čunát | (I do that at work.) | 17:48:37 |
| edef | but a bloom filter only kicks the can down the road, it only answers "probably present", etc | 17:49:30 |
| edef | you'll improve p95 latency but not p100 | 17:49:49 |
| edef | it is an improvement, but the actual hashset fits in memory easily | 17:50:28 |
| Arian | Our p100 times probably pretty bad now too. S3 fetch across the continent is not particularly fast | 17:51:18 |
| edef | this is true, yes | 17:51:32 |
| edef | sticking the narinfos in postgres would be easy and get us a bloom filter with a CREATE INDEX | 17:52:13 |
| edef | i forget what backends we currently have for the snix narinfo service but "just a database" is the straightforward answer | 17:52:59 |
| edef | an object store full of text files is just an awkward answer | 17:53:16 |
| Arian | What if we only do my idea for Nars and not narinfos | 17:53:36 |
