Can you collect some perf result on that? Thanks!

I also found this regression on debug mode and Sysbench QPS dropped to almost zero. However everything works well on release mode.

single-binary doesn't work well, while multi-binary looks good. Is there anything that might go wrong with the madsim update? cc @wangrunji0408 there should be no problem with madsim

Probably not, seems to be a problem with frontend node.

And ok_or_else seems to consume significant amount of CPU time.

... because backtrace is slow to create

In multi-binary:

backtrace captured in 1ms
backtrace captured in 0ms
backtrace captured in 0ms
backtrace captured in 0ms
backtrace captured in 0ms
backtrace captured in 0ms
backtrace captured in 0ms
backtrace captured in 0ms
backtrace captured in 0ms
backtrace captured in 0ms

In single-binary:

backtrace captured in 393ms
backtrace captured in 369ms
backtrace captured in 370ms
backtrace captured in 365ms
backtrace captured in 362ms

Probably a bug with the Rust compiler, but I would still recommend not returning so many errors.

Related function:

    pub fn get_schema_by_name(&self, db_name: &str, schema_name: &str) -> Result<&SchemaCatalog> {
        self.get_database_by_name(db_name)?
            .get_schema_by_name(schema_name)
            .ok_or_else(|| {
                let x = Instant::now();
                let err = CatalogError::NotFound("schema", schema_name.to_string()).into();
                println!("backtrace captured in {}ms", x.elapsed().as_millis());
                err
            })
    }

The suspicious thing is the dwarf region in the binary.

On 2022-07-29:

objdump target/debug/risingwave  --dwarf=frame | wc -l
     371

On 2022-10-16:

objdump target/debug/risingwave  --dwarf=frame | wc -l
 14671494

Shall we file a bug report to rust-lang? (Though I believe this is more of an LLVM bug / LLVM upgrade)

Maybe related: rust-lang/rust#98051

I would suggest keep this issue open while changing frontend to return Option instead of error.

I would suggest keep this issue open while changing frontend to return Option instead of error.

I think most of the errors generated in the frontend are for responding to the user, which cannot be avoided.

Is it caused by capturing the stacktrace?

I think most of the errors generated in the frontend are for responding to the user, which cannot be avoided.

No. The e2e test cases that timed out should not have any user-facing error. All tables exist.

Is it caused by capturing the stacktrace?

Yes.

There seems to be some new features in Rust / LLVM that generates a special kind of debug info DWARF, which contains stack unwinding region information. In debug mode it's just slow to rewind.

It's not a big issue but still affected the developer's experience 🥲 What is our solution?

1. Wait for rust community to fix it. (Don't know how long it will take)
2. Refactor the code that captures the stack-trace into error
3. Any other ideas?

Wait for rust community to fix it.

I don't think that's something to be fixed. It looks more like a feature that provides better backtrace.

Refactor the code that captures the stack-trace into error

I think a fix is going on to change them to cheaper CatalogEror.

cc. @BowenXiao1999

I've opened an issue for this -> #6157.

Backtrace capturing is not only slower, but can also lead to a segfault: #6205

The performance problem is much better now. It took me 200s to run slt-all for multi-binary and 250s for single-binary (when the issue opened it was forever).

Closing this issue in favor of #4077 and #6205.

The slowness is acknowledged by the doc of Backtrace::capture, and our take on it would be to avoid unnecessary backtrace capture via #4077.

The segfault is worse and tracked by #6205.

TImeout is a separate issue. reopen.