Struggle with a tools
As a slopocoder, I run 10 parallel sessions this year, and some amount of my work now is to maintain my forks of the ecosystem.
It all starts with the fact that each agent session spins up its own full rust-analyzer: 5+ GB each (on my workspace a full RA context is 15–25 GB). So I did MCP work (rust-code-mcp #12, my fork is enomado/rust-code-mcp) so that there is one shared daemon instead of a server per session. There are tons of such projects, but officially none of this is recognized (rust-analyzer #22765). My slop contributions are rejected (rust-analyzer #22737, #23288), this guy's (alex09x) contributions are rejected (rust-analyzer #23454, in-memory worktree sharing), and I'm banned on Rust GitHub (surprisingly, in all organizations, because of RA).
Then I realized that RA simply leaks. Well, technically it never lets go: the LRU capacity knobs have been no-op stubs since the salsa migration (FIXME(salsa-transition)), so nothing is ever evicted. Gigabytes sit in macro expansions and parse trees, not in type inference. I did some patches in my salsa and RA forks: runtime LRU capacities by query name, per-query heap accounting. And this one: salsa #1304. Did GC. Optimize peak ram consumption. It works somehow and I'm happy.
After that I wrote GC for the target folder, because I'm tired of doing cargo clean each hour, because it grows to 400 GB.
Today I realized I want parallel builds, and another shit happened: this flag (-Zfine-grain-locking) simply doesn't work because of deadlocks (cargo #17508). Two cargos building the same dirty units wait on each other forever; in my repro, two cargo check from scratch deadlock 7 times out of 8. For now I run a small watchdog that reads /proc/locks and kills the youngest cargo in a cycle. No, it's awesome that the team realized they are already want this feature and are taking some measures, but according to the issue, the story is very slow, and there is another and another initiative (cargo #4282 open since 2017, #5931, #16147). Should I bother to fork cargo?
I'm curious: am I the only alone with a 2M+ LOC codebase, or does some good practice actually exist? Or does any Rust dev have 1 TB of RAM and a 40 TB SSD?
And a question with an asterisk: when will the Rust project actually start making the ecosystem usable? Here is the official plan for cargo build to stop recompiling everything after cargo check (Rust project goals 2026, "incremental system rethought", see Funding): 12 months, $75k–120k, "partially funded", and months 1–3 include "establish formal communication with the cargo and compiler teams". Meanwhile the same money in tokens would close half the issue tracker in YOLO mode, with the PRs then auto-closed by the AI policy, of course.
This post was auto-deleted in reddit and I dont know the reason and this platform does not allow me to post urls. I dont know any channel to discuss this stuff but I feel very fatigued from this situation. It continues for months and it seems like nothing changes.
1 post - 1 participant
原文:https://www.a8q9.cn/en/article/struggle-with-a-tools-faacec.html