You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Agent test an_app_that_misbehaves_is_refused_or_dropped_and_the_rest_go_on needs more than 16 GiB #15
On unmodified main (98bbab3), one agent test uses more than 16 GiB of memory on Windows. That's enough to make cargo xtask ci fail on a machine with a 16 GiB limit, and it might point at an unbounded queue in the agent.
What I see
All runs were on Windows 11 with Rust 1.99.0, inside a job object that caps the whole process tree at 16 GiB (commit memory).
aborts: memory allocation of 786432 bytes failed, exit 0xc0000409
16.0 GiB
cargo test -p craft-cloud-agent --test live_documents --locked -- --test-threads=1
aborts in the same test
16.0 GiB
0xc0000409 is how Rust's allocation-failure abort shows up on Windows; it isn't a stack overflow. The first run also compiled the test binary, so its peak includes that; the other two ran on the finished build.
So the test alone, with one test thread, keeps allocating until the cap. I mentioned a_craft_save_goes_up_with_its_record in #13 and #14; that was wrong, since that test only happened to be the one still running. Without a cap the test may simply pass on a machine with more memory, and I haven't measured how far it goes.
Where it might come from
From reading the code, not proven:
The test sends 60 updates of 200,000 characters each (apps/craft-cloud-agent/tests/live_documents.rs:653). That's 12 million characters retained in the Loro history, plus replicas and exports.
The agent's server-output channel is unbounded (apps/craft-cloud-agent/src/live/hub.rs:1373). The 10-second send timeout limits how long a stalled send waits, not how many bytes queue up.
The app's event channel is an unbounded std::sync::mpsc::channel (crates/app/src/live.rs:365).
If it's one of the queues, a local app that misbehaves could drive the agent's memory without limit, which seems worth a bound outside the test too. If it's only the Loro history of 12 M characters, a smaller flood in the test would do.
Repro
Run the second command above on a machine or VM with about 16 GiB, or under a memory cap. I'm happy to dig further, with a heap profile or a bounded-queue experiment, if that's useful to you.
Activity