As I mentioned in an earlier post, me and other contributors are porting Newsboat to Rust. We’re working on utility functions right now, but soon, it will be time to port more substantial chunks. To prepare, I ported our logger class. Here’s how that went.
The proper way to do logging in Rust is to add
log to your Cargo.toml and be done with it. However, this couldn’t work for Newsboat: we have a custom log level called “user error”, messages on which are not only written to the main log, but are also copied to a special one. The
log crate doesn’t allow for custom log levels, so we had to roll our own.
My plan was simple: write a single-threaded logger, put it behind a
Mutex, done. I had to immediately correct this when I learned that I can’t create a
static with a
lazy_static! solved that. Then it took me about a week to realize that this design will needlessly serialize the threads, as they’ll have to acquire the mutex just to check the current log level. Back to the drawing board, then.
A better design, lifted from the
log crate, is to put the current log level into an atomic, and protect the rest with the mutex. That way the checks are relatively cheap and concurrent, while actual logging is still serialized properly.
Even with that new design, my Rust logger was still 1.5 times slower than the old C++ one. That’s when I really started making mistakes instead of progress. I read up on the internals of
Once, and other primitives I was using. I tried varying the design in hopes of making the log level check even faster. I did learn about Rust a bit, and I sped up the timestamp formatting on the Rust side, but I still couldn’t make the new logger go as fast as the old one.
And then it hit me. All the code was in Rust now, the only thing left on the C++ side was a thin wrapper over the C API that Rust provided. That wrapper was a tad too thin, because it didn’t check the current log level. All it did was format the log message and call Rust, which would discard the message immediately if the current log level wasn’t high enough. In the worst case— i.e. with the logging disabled—C++ wasted a ton of time formatting messages.
Once I started checking the level on the C++ side too, the performance of the new logger became indistinguishable from that of the old one. Yay!
I know I promised you a rant about Rust’s unit testing framework, but it didn’t fit well with the tale of the logger. Stay tuned if you don’t want to miss it when I finally post it ;)
Drop me a line! (wonder where’s the comments form?)