2 months ago
Hi,
I'm on the Hobby plan running a Python image-processing service, and I'm seeing large performance variance between deployments that I'd like to understand before deciding whether to upgrade.
The pattern: a given container is either fast or slow for its entire lifetime, and it flips at the next redeploy. Same image digest, same code, no config changes. Across roughly 25 containers over five days, about four were consistently ~3x slower than the rest on identical workloads. A slow container stays slow until it's replaced, and the replacement is usually fast again with no intervention.
The part that makes me think it's host-level rather than resource limits: during slow periods the service uses fewer total CPU-seconds than during fast ones, while wall-clock time triples. The work appears to lose parallelism — operations that normally spread across several cores seem to run on about one. The container never approaches its own CPU or memory limits in either case (peaks around 1.3 vCPU against 8 available, and about 500MB of RAM).
My questions:
-
Does host placement or CPU allocation differ between Hobby and Pro? I'm trying to work out whether Pro would reduce this variance, or whether it only raises limits I'm not currently hitting.
-
Is there anything about how CPU is allocated on shared infrastructure that would explain a workload losing effective parallelism on some hosts but not others?
-
Is dedicated compute available below the Enterprise tier, or is that the only option for consistent performance?
Thanks,
John
5 Replies
2 months ago
This thread has been opened as a bounty so the community can help solve it.
Status changed to Open Railway • about 2 months ago
2 months ago
im also on hobby, this looks like cpu contention on a shared host, not your limits, since you're nowhere near your ceiling.
-
hobby shares cpu with other tenants via cgroup quotas, pro might get better priority but not sure it removes contention fully, staff would know .
-
yes, contention causes exactly this, your threads wait their turn on a busy host so work loses parallelism, thats why cpu-seconds drops while wall clock triples
-
not sure on dedicated compute below enterprise, worth asking staff directly
manuproject
im also on hobby, this looks like cpu contention on a shared host, not your limits, since you're nowhere near your ceiling. 1. hobby shares cpu with other tenants via cgroup quotas, pro might get better priority but not sure it removes contention fully, staff would know . 2. yes, contention causes exactly this, your threads wait their turn on a busy host so work loses parallelism, thats why cpu-seconds drops while wall clock triples 3. not sure on dedicated compute below enterprise, worth asking staff directly
2 months ago
Thanks for the response. I would love to ask staff but it seems like my only option is to upgrade to Pro first in order to speak with them. I guess it may be worth it.
theoakagency
Thanks for the response. I would love to ask staff but it seems like my only option is to upgrade to Pro first in order to speak with them. I guess it may be worth it.
2 months ago
yes you can its just 5$ to upgrade to HOBBY
an its going to worth it
2 months ago
I’m actually dealing with a similar question with a backend I’m currently running on Railway, so this is really interesting to me.
I’d also like to know whether upgrading from Hobby to Pro would actually change CPU scheduling/host placement or reduce contention, rather than simply increasing the resource limits. If the workload isn’t reaching the CPU or memory limits in the first place, it would be useful to know whether Pro provides any meaningful performance or isolation benefit.
Would be great if someone from Railway could clarify whether this kind of deployment-to-deployment variance is expected on shared infrastructure, and what the options are for getting more consistent CPU performance without going all the way to Enterprise.
manuproject
yes you can its just 5$ to upgrade to HOBBY an its going to worth it
2 months ago
Hello. I'm actually paying $5 for Hobby and I see this when I try to reach out to Railway. Are you aware of another way to reach them directly?
Thank you!
Attachments