5 ms·

How much more performance do you need to get from Rust over Python (even Cython, PyPy, Numba, etc) to justify the extra development cost? A 2x gain is certainl

How much more performance do you need to get from Rust over Python (even Cython, PyPy, Numba, etc) to justify the extra development cost? A 2x gain is certainly not worth it. A 10x gain? Maybe. But that is hard to achieve when much of your "compute" is spent on the DB side of things. How many startups actually scale out of needing a few non-db instances?

soggybutter 4y ago
My team helps run and deploy a python service that is entirely CPU bound. It accepts an input, performs some computation, and returns a result without any sort of I/O outside of the initiating HTTP request. In the past week it's averaged around 144 req/s with a p95 latency of ~1s. We average ~80 "instances" to maintain this level of performance. I have very little doubt that, if given the opportunity to rewrite this in rust, we could smash 10x perf improvements. Could we also get more perf out of tuning our python code better? Definitely. Do I think there's 10-20x improvement waiting to be uncovered? No. Unfortunately (fortunately?) we're at a stage that it makes more sense to throw ludicrous sums of money at it than it does to ground up rewrite.
adastra22 4y ago
Rust easily gets 10x improvement in performance over Python in a lot of applications. This is absolutely my experience. A better statement is that it doesn’t get 10x improvement over Go, or other ergonomic compiled languages.
- steveklabnik 4y ago
  Also in the context of cloud spend: way way way less RAM, which can translate directly to dollars.
  - cozzyd 4y ago
    Yeah, I just wrote a simple daemon in C that had no performance requirements ( listening to a udp socket and dumping stuff into a pgsql db once a minute) because the Python program would use like at least 20x the RAM. When you're running a bunch of things on a resource constrained place (e.g. a single computer that has to do a ton of things sitting on a rack on the Greenland ice sheet), even just the base Python memory usage from a new process adds up...
ReflectedImage 4y ago
It will be 25x gain on average and your development costs will triple. Is that worth it? Depends on what you are building....

soggybutter 4y ago

My team helps run and deploy a python service that is entirely CPU bound. It accepts an input, performs some computation, and returns a result without any sort of I/O outside of the initiating HTTP request. In the past week it's averaged around 144 req/s with a p95 latency of ~1s. We average ~80 "instances" to maintain this level of performance. I have very little doubt that, if given the opportunity to rewrite this in rust, we could smash 10x perf improvements. Could we also get more perf out of tuning our python code better? Definitely. Do I think there's 10-20x improvement waiting to be uncovered? No. Unfortunately (fortunately?) we're at a stage that it makes more sense to throw ludicrous sums of money at it than it does to ground up rewrite.

adastra22 4y ago

Rust easily gets 10x improvement in performance over Python in a lot of applications. This is absolutely my experience. A better statement is that it doesn’t get 10x improvement over Go, or other ergonomic compiled languages.

steveklabnik 4y ago
Also in the context of cloud spend: way way way less RAM, which can translate directly to dollars.
- cozzyd 4y ago
  Yeah, I just wrote a simple daemon in C that had no performance requirements ( listening to a udp socket and dumping stuff into a pgsql db once a minute) because the Python program would use like at least 20x the RAM. When you're running a bunch of things on a resource constrained place (e.g. a single computer that has to do a ton of things sitting on a rack on the Greenland ice sheet), even just the base Python memory usage from a new process adds up...

ReflectedImage 4y ago

It will be 25x gain on average and your development costs will triple. Is that worth it? Depends on what you are building....