| 20 May 2026 |
 K900 | I'm not sure what kind of optimization you're looking at here, but generally GC or any kind of store queries aren't the bottleneck | 12:07:33 |
 Cobalt | The main optimizations were how to reduce the number/cost of queries for evaluating the subset size required when looking scheduling a set of builds. They might not be a bottlneck by themselves but caching and/or applicability of a stochastic data structure seems an interesting extension. My supervisor was interested in this specific sub-problem as it relates a bit to his own research iirc. | 12:13:23 |
| Cobalt | fast-nix-gc does not really have anything related to this, it just mentions that they load the paths into a graph for the GC search first instead of querying the store for all lookups. | 12:14:15 |
| K900 | Well right now the scheduling is very stupid | 12:14:20 |
| K900 | There's no locality awareness | 12:14:30 |
| K900 | Or hell job size awareness | 12:14:32 |
| K900 | Improving it will definitely help a little, but the big bottleneck is still the coordinator itself | 12:15:32 |
 hexa | the scheduler | 12:15:50 |
| K900 | The coordinator as in the machine | 12:16:01 |
| K900 | But yeah | 12:16:02 |
| hexa | the coordinator is the process that runs the remote build | 12:16:14 |
| Cobalt | So just to understand this a bit more, a significant problem is the performance of the software running the scheduler/coordinator (so the queue runner)? | 12:17:07 |
| K900 | It's not even the software necessarily | 12:18:34 |
| K900 | It's the design of the whole thing that requires a lot of copying data around | 12:18:44 |
| Cobalt | A montivation of the optimizations was to ensure that scheduling was supposed to stay cheap-ish so I would try to not compromise this too much. | 12:18:53 |
| K900 | And also the fact that everything is xz compressed in transport which is extremely overhead | 12:18:59 |
| Cobalt | Is that regarding the data exchange of the build outputs, RPC or artifacts (logs)? | 12:19:47 |
| K900 | Build outputs | 12:20:05 |
| K900 | Logs are negligible by comparison | 12:20:12 |
| K900 | And RPC is just normal Nix daemon protocol over SSH | 12:20:18 |
| Cobalt | Maybe I misunderstood something there but aren't they uploaded directly from the builder to S3 bucket? | 12:20:37 |
| K900 | They are not | 12:20:44 |
| Cobalt | * Maybe I misunderstood something there but aren't they uploaded directly from the builder to the S3 bucket? | 12:20:45 |
| K900 | They are currently copied to the coordinator for signing | 12:20:52 |
| Cobalt | Oh, well, that does sound really expensive for bandwidth (and compute with compression). | 12:21:27 |
| Cobalt | * Oh, that does sound really expensive for bandwidth (and compute with compression). | 12:21:33 |
| K900 | It is | 12:21:47 |
 Vladimír Čunát | And S3 is over the Atlantic. | 12:21:53 |
| Vladimír Čunát | There are lots of "design problems". | 12:22:09 |
| Cobalt | Interesting, I will have to take a closer look at it then (especially if my knowledge of the architecture is apparently inaccurate).
I will still probably stick with scheduling for now due to other constraints on the topic but thank you for the extra information. | 12:24:03 |
