Self-hosting Nextcloud + immich + matrix = every month. 1TB plan from OVHcloud, keep 3 days in.
Tipo I know thumbnails, sync, federation... mas how do people actually stay under? Not "use less"—I mean accounting tricks, compression, cache layers, anything. What am I blind to?
Running Debian, nginx, nothing exotic. Backup to Vultr cold storage. Still bleeding.
Member
Virtualization tax on bandwidth accounting is real—some providers count ingress + egress separately, others only egress. Check your control panel's definition. OpenVZ nodes especially lie about interface counters; KVM is trustworthy here.
For your stack: immich generates thumbnails on upload. Disable remote machine learning if enabled—those pings add up. Matrix federation is worst offender; use conduit instead of synapse, lower overhead.
What's your backup method?
You need to check your backup script.
I had same problem, was using zfs send to Vultr but didnt set -I flag so full stream every night 300GB daily.
After fix with incremental only 2GB.
Also compress with zstd cant believe the difference.
Prices for bandwidth overage crazy $0.05/GB!
Wesh @kate3 le server حلو but accounting tricks incha'allah only get you so far
My fix : nginx cache layer for immich, 7 days TTL. Reduces repeat thumbnail fetch by ~60 %. I use https://nginx.org for that. Also : disable Nextcloud client preview generation, do it server-side once
French spacing here ; check your matrix media retention , default is forever !
Member
Thanks for the detailed context, emalite—this is a common pain point we see at Contabo too.
A few things to check via your OVHcloud panel or DM me:
• Whether "1TB" is combined or egress-only
• If private traffic to Vultr counts against plan (some providers exempt object storage partners, some don't)
• Whether your nginx logs show unexpected crawler traffic
Quick wins:
— Enable brotli compression at nginx level
— Set immich to generate WEBP thumbs, not JPEG
— Matrix: set media_retention to 30 days max
— Jane @ Contabo
1TB is nothing. My old Supermicro X9 at colo pushes 4TB monthly just from rsync mirrors, no users.
Real talk: check if your "cold storage" backup is actually cold. Some APIs poll metadata constantly. Had a client burning 200GB/month on S3 ListObjectV2 calls alone.
Also: what generation NIC? I350-T4 vs newer X550 changes checksum offload behavior, affects measured throughput vs actual wire bytes. Not huge but at your margin, matters.
If you're serious, ship a 1U Atom box to Leaseweb Utah — https://www.leaseweb.com. Flat €30 for 10TB. No games.
What NIC generation? I350-T4 vs X550 matters for your offload
Member
Same issue with zfs send bro, full stream killed my quota
Member
Conduit instead of synapse (really?) (the overhead difference) (enough to matter?) (at 1TB?)
25 days, but this thread's too useful to let sink.
This is the real answer. Most people bleeding quota aren't even looking at their backup patterns. zfs send without flags, rsync without --link-dest, whatever—it's always the cron job you wrote six months ago and forgot.
For the original stack: immich has a "generate thumbnails" toggle that's client-side now. If you're still doing server-side, switch it. Saves a surprising amount of egress on mobile sync.
Matrix federation though—honestly, just block the big public rooms. Your homeserver doesn't need to know about libera.chat memes.
Anyone ever try Varnish in front of nginx for the Nextcloud web UI? Curious if the cache hit rate is worth the RAM tradeoff at this tier.
