Is RwTxn Send?

[elbaro]

I found that multi-threaded tokio runtime moves the task between threads, and this leads to let tx = env.write_txn() and tx.commit() to run in different threads. And the next call to env.write_txn() hangs.

let tx = env.write_txn()?; // thread A

while let Some(row) = stream.next().await {
    insert_row(&mut tx, row)?;
}

tx.commit()?;  // thread B

let tx = env.write_txn()?;  // hangs

[indexing] IndexBuilder tx is created in thread=ThreadId(9)
████████████████████████████████████████ 4317834/4317834
[indexing] IndexBuilder tx is committed in thread=ThreadId(17)
[task] lmdb - getting WriteTx in thread=ThreadId(17)  # hangs

Switching to a single-threaded tokio runtime did not have a deadlock.

I found this in jnwatson/py-lmdb#89 (comment) :

A write Transaction can only be used from the thread it was created on.

Is it okay to commit a write transaction in a different thread?
If not, what is a recommended approach to interleave async workload with transactions?

[Kerollmops]

Hey @elbaro 👋

Thank you for the report. It's indeed a problematic issue. We are using heed for a long time at Meilisearch and probably a lot in async multithreaded context and never had the issue.

However, it seems to be true that a write transaction MUST be committed on the original thread. Maybe, we did that by mistake.

We even have tests to check that we can move write transactions between threads 🤦‍♂️