Sleep in Async Contexts

  1. More CPUs, please
  2. Shipping it
  3. The investigation
  4. The problem
  5. Async mindfield
  6. Take aways

I recently built a Rust library that allows you to query Impala databases for an API.

The library interfaces with an Impala database via the Thrift interface, namely via the interface defined by the TCLIService.thrift definition file. Using the Apache Thrift CLI tool, the bindings file for Rust can be generated to be used with the thrift crate.

thrift -out src --gen rs -r ./thrift/TCLIService.thrift

The synchronous API was first built to test things out (which involves a lot of transformations like transposing rows, getting metadata, and more shenanigans). Then the asynchronous API was built with the help of bb8 and thrift-pool which allows the underlying TCP connections to be reused since they are expensive to reopen every time a query fires.

Since this was the first time I was touching code related to such low-level interfaces, there were some bugs like a query only returning a maximum of 1024 rows due to a limit imposed on the server side, requiring the query to be sent again on the same session with all the same parameters.

Nonetheless, we deployed the first iteration of the product which uses this Impala libary and I noticed something weird.

More CPUs, please

Since Impala iteself is kinda slow, I expected that the requests (that requires calling to Impala) to be slow. My response was to bump up the number of CPUs, since at the time of the pre-release, there was only 1 CPU instance allocated to the whole container running the server.

The reasoning was: if there's more CPUs, then the actix-web server can use more CPUs to run tasks, and there will be less contention between the system processes and the server.

After deploying the new version onto Google Cloud, I tested the application by requesting some routes, and everything seemed fine to me.

Shipping it

On the day of release, multiple users reported that there was issues loading resources from the API. They were being timed out constantly.

Looking at the logs, I found that some routes on the server which were not supposed to be slow, were slow.

An example is the /auth route. This route checks if your auth cookie is still valid, and it should be nearly instant (!), but it was taking more than 4 seconds, and sometimes even more than 10 seconds to be processed. When this happens, it triggers Google Cloud to timeout the request and return a HTTP 408, causing cascading errors to the users.

The investigation

I tried to reproduce the issue on my local machine by limiting the server to only utilise 1 thread, firing requests to the slow endpoint, and then firing multiple requests to the fast endpoint in quick succession.

As expected (not to my advantage), the supposedly fast requests to /auth were being slow, and they were suspiciously slow because the response times were almost the same as the slow request on /chart!

image showing slow requests to /auth and /chart

The problem

As the title suggests, the problem is caused by the usage of std::thread::sleep in an async context.
async fn wait_to_finish(
    &mut self,
    op: &TOperationHandle,
    start: &Instant,
) -> thrift::Result<()> {
    // ...
    let sleep_duration = get_sleep_duration(start);
    log::trace!(" sleeping for {:.2?}", sleep_duration);
    // Oops!
    // ...

I promptly switched the implementation to use tokio::time::sleep instead, which fixed the issue:

image showing after fixes of faster /auth route

The documentation of std::thread::sleep says:

Puts the current thread to sleep for at least the specified amount of time.

The thread may sleep longer than the duration specified due to scheduling specifics or platform-dependent functionality. It will never sleep less.

This function is blocking, and should not be used in async functions.

During the porting of the synchronous API to the asyncronous API of the Impala library, I didn't manage to catch this and mostly ported everything 1-to-1 (since the mechanism of firing queries is the same). I thought to myself: isn't there a clippy lint for this that could have caught this issue?

Turns out, there isn't.

There was an issue opened in rust-clippy titled "Lint idea: find known-blocking constructs in async contexts", but it wasn't in favour to be implemented in clippy because it requires global analysis, which is something that clippy shouldn't do due to perfomance reasons.

Async mindfield

All of this reminds me of how difficult it is to implement async Rust. I remembered one time where I freaked out when I saw the Future trait:
pub trait Future {
    type Output;
    // Required method
    fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output>;

Pin? Poll? What the hell is this? Let me see the definitions... oh, okay nevermind.

Suffice to say I quietly stepped away from this, and I just never return Future-y things from functions or even try to touch Future related things by using async fn and .await everywhere that I could. Maybe one day I am able to muster the confidence and invest time to truly understand the underlying mechanism of Futures.

Don't get me wrong, I am quite satisfied with how Rust handles async and being extremely explicit, but I do wish there was some tool that would have caught this issue for me, which would have in turn make working with async easier.

Take aways

Some key take-aways: