a month ago
Hello,
We run a CPU-heavy audio-mastering worker (Celery, Python and ffmpeg, 8 vCPU / 8 GB). The same work takes 2 to 4 times longer depending on which host a deployment lands on, and we need predictable performance before launch.
What we see: the same task on the same 4:10 audio file renders in 11 to 18 s on some deployments and 33 to 54 s on others (up to 141 s for a heavier variant). Every deployment whose start-up banner shows kernel 6.12.12+bpo has been slow (39 of 39 renders). Most deployments showing 6.18.15+deb13 are fast, but some had slow spells on 02 to 04 October. The slow renders ran with nothing else on the worker; CPU peaked around 3.3 vCPU and memory around 3.6 GB, so limits were not reached.
Slow periods (UTC): 02-10 around 15:15; 05-10 16:27 to 06-10 11:18; 06-10 13:10 to 16:47. Fast periods: 06-10 11:18 to 13:10, and from 06-10 16:47 onward.
Questions: (a) Were the slow deployments on shared hosts with contention (noisy neighbour), or on a different hardware generation? (b) Can a service be placed on a consistent host pool (for example Metal Gen-2) so CPU performance is predictable? (c) Is there a plan or setting that guarantees dedicated or stable CPU for a worker like this?
Happy to share project, service and deployment IDs privately with Railway staff.
Thank you,
Michael
1 Reply
a month ago
Replicas in the same region can run on different hardware, so the same vCPU and memory limits don't always produce the same CPU throughput. That's consistent with what you're measuring, but we haven't confirmed which hardware your slow deployments ran on, or whether contention played a part, so we can't attribute those specific runs to either cause.
On (b) and (c): there's currently no setting or plan that lets a service choose a CPU model, pin itself to a hardware generation or host pool, or guarantee dedicated CPU. Picking a region doesn't guarantee this either. A redeploy can change placement and sometimes improves performance, but it can also land on slower hardware again, so it isn't a dependable fix.
For launch planning, the most reliable approach is to size your capacity around the slower render times you've observed, not the fast ones. Any contractual isolation or placement requirement would have to be reviewed case by case and isn't guaranteed by a plan upgrade.
Status changed to Awaiting User Response Railway • about 2 months ago
a month ago
This thread has been marked as solved automatically due to a lack of recent activity. Please re-open this thread or create a new one if you require further assistance. Thank you!
Status changed to Solved Railway • about 1 month ago