On this page (8)
  1. Idle and peak RAM, app by app
  2. What idle and peak mean in our lab
  3. The heaviest apps
  4. The light ones
  5. How to add up a stack
  6. What the numbers do not cover
  7. Download the data
  8. Related

In our lab most self-hosted apps fit in 1 GB of RAM by our sizing rule, and Immich is the exception. Vaultwarden, Home Assistant, Jellyfin and Nextcloud all come to that figure, while Immich needs 8 GB with machine learning on and 4 GB with it off. Idle use is lower than any of these: Pi-hole idled at 9 MiB and Immich at 1.3 GiB. The table gives idle and peak RAM, the load behind the peak and the guide for every self-hosted service we have measured, all on the same kind of machine.

Idle and peak RAM, app by app

The apps are sorted by idle RAM, heaviest first. “RAM by our rule” is the memory to plan for: the highest reading in any run, times 1.5, rounded up to the next of 1, 2, 4, 8, 16 or 32 GB (the full rule is on the methodology page). It is not the idle figure, and it is not the median peak either.

Idle and peak RAM of every published self-hosted app in the lab, median of the runs with the range, the load behind the peak, the RAM our sizing rule recommends, the runner CPU and the guide
AppVersion, testedIdle RAM (MiB)Peak RAM (MiB)What the peak load wasRAM by our ruleRunner CPUGuide and data
Immich (ML job concurrency 1)v3.3.5
2026-10-28
1,407
(1,335 to 1,419)
3,685
(3,613 to 3,811)
20 photo uploads and every background job, machine learning one job at a time8 GBAMD EPYC 9V45Immich RAM Requirements
result JSON
Immich (4 GiB cap)v3.3.5
2026-10-28
1,406
(1,342 to 1,411)
3,073
(2,917 to 3,298)
20 photo uploads and every background job, containers capped at 4 GiB in total8 GBAMD EPYC 7763Immich RAM Requirements
result JSON
Immichv3.3.5
2026-10-28
1,321
(1,308 to 1,330)
3,701
(3,538 to 3,802)
20 photo uploads, then every background job (machine learning on)8 GBAMD EPYC 9V45Immich Docker Compose
result JSON
Immich (no machine learning)v3.3.5
2026-10-28
1,090
(1,075 to 1,132)
1,671
(1,633 to 1,701)
20 photo uploads and every background job, machine learning off4 GBAMD EPYC 7763Immich RAM Requirements
result JSON
authentik2026.8.5
2026-11-08
1,067
(840 to 1,071)
1,109
(918 to 1,113)
60 password logins, then 3,000 authorization checks2 GBIntel Xeon 6975P-Cauthentik Docker Compose
result JSON
Immich (no machine learning, 2 GiB cap)v3.3.5
2026-10-28
1,063
(1,037 to 1,324)
1,536
(1,501 to 1,644)
20 photo uploads and every background job, machine learning off, capped at 2 GiB in total4 GBIntel Xeon Platinum 8370CImmich RAM Requirements
result JSON
Seafile13.0.28
2026-10-16
616
(590 to 759)
905
(876 to 911)
one user: upload, list and download 103 files (about 200 MB) through the web API2 GBAMD EPYC 9V45Nextcloud vs ownCloud vs Seafile
result JSON
Home Assistant2026.10.4
2026-10-28
363
(354 to 366)
368
(358 to 369)
eight automations (six on timers of 2 to 7 s), plus one WebSocket client for 120 s1 GBAMD EPYC 7763Home Assistant Docker Compose
result JSON
Nextcloud35.0.2
2026-10-18
297
(295 to 301)
452
(402 to 479)
one user: upload, list and download 103 files (about 200 MB) over WebDAV1 GBAMD EPYC 7763Nextcloud Docker Compose with Postgres, Redis and Cron
result JSON
Jellyfin12.4
2026-11-03
186
(184 to 187)
406
(405 to 406)
library scan of 13 generated films, then one software transcode for 60 s1 GBAMD EPYC 7763Jellyfin Docker Compose
result JSON
ownCloud Infinite Scale8.2.2
2026-10-28
152
(95 to 165)
varies between runs
314
(313 to 315)
one user: upload, list and download 103 files (about 200 MB) over WebDAV1 GBAMD EPYC 7763Nextcloud vs ownCloud vs Seafile
result JSON
Uptime Kuma2.5.9
2026-11-03
123
(122 to 124)
139
(138 to 140)
30 monitors checked every 20 s for 10 minutes1 GBAMD EPYC 7763Uptime Kuma Docker Compose
result JSON
Authelia4.39.28
2026-11-08
56
(48 to 65)
607
(518 to 650)
60 password logins, then 3,000 authorization checks1 GBIntel Xeon 6975P-CAuthentik vs Authelia
result JSON
AdGuard Homev0.107.79
2026-11-06
2652DNS queries at up to 1,000 per second, 900 names1 GBAMD EPYC 9V74AdGuard Home vs Pi-hole v6
result JSON
Vaultwarden with Caddy1.37.5
2026-11-03
22
(21 to 22)
67
(66 to 69)
60 s of API reads and writes on a 100-item vault, over HTTPS through Caddy1 GBAMD EPYC 7763Vaultwarden Docker Compose
result JSON
FileBrowser Quantumv2.0.2-stable
2026-10-28
21
(20 to 22)
36
(34 to 39)
one user: upload, list and download 103 files (about 200 MB) through the HTTP API1 GBIntel Xeon Platinum 8573CSelf-Hosted Alternative to Google Drive
result JSON
File Browser2.63.23
2026-10-16
15
(13 to 15)
23
(21 to 23)
one user: upload, list and download 103 files (about 200 MB) through the HTTP API1 GBAMD EPYC 9V45Self-Hosted Alternative to Google Drive
result JSON
Pi-hole v62026.10.1
2026-11-06
9
(9 to 10)
18DNS queries at up to 1,000 per second, 900 names1 GBAMD EPYC 9V74Pi-hole Docker Compose (v6)
result JSON

18 rows. On a narrow screen the table scrolls sideways and the app names stay in place. Each figure is the median of 3 runs, with the lowest and highest run in brackets where they differ after rounding. Tested dates run from 2026-10-16 to 2026-11-08. Every run was on one machine type: github-actions ubuntu-latest (2 vCPU, 8 GB). Left out because they run a command-line tool or a backup job instead of a service, so their container figures do not describe the tool: Vaultwarden backup and restore, restic backup benchmark, BorgBackup backup benchmark, Kopia backup benchmark.

The CPU was not the same on every row. The same runner type lands on different physical CPUs, and all the runs of each row ran on one model. RAM carries over to other machines far better than speed does, but a row can differ from the next by the CPU as well as by the app.

  • AMD EPYC 7763: Immich (4 GiB cap), Immich (no machine learning), Home Assistant, Nextcloud, Jellyfin, ownCloud Infinite Scale, Uptime Kuma, Vaultwarden with Caddy
  • AMD EPYC 9V45: Immich (ML job concurrency 1), Immich, Seafile, File Browser
  • AMD EPYC 9V74: AdGuard Home, Pi-hole v6
  • Intel Xeon 6975P-C: authentik, Authelia
  • Intel Xeon Platinum 8370C: Immich (no machine learning, 2 GiB cap)
  • Intel Xeon Platinum 8573C: FileBrowser Quantum

Two things to read before you compare rows. The loads differ, so the peaks answer different questions: one photo import, one user moving 103 files, 30 monitors, 60 logins. And a row is marked “varies between runs” when its highest run was more than 1.5 times its lowest, so read it together with the range under the median.

What idle and peak mean in our lab

Every number comes from one harness. It brings an app up with Docker Compose from nothing, runs the setup steps, waits until the total memory of all its containers is flat, and only then measures. The full rule is on the methodology page; the short version:

  • Idle is the mean memory of the whole stack over a quiet window after it has settled. Some apps use more for a few minutes after start-up (Immich’s server does), and a fixed wait would call that idle.
  • Peak is the highest total memory of all the containers at the same moment during the load step in the table: a scripted client doing the thing named in the “What the peak load was” column. It is a synthetic, single-user load, not your traffic.
  • Memory is what docker stats reports for each container: usage minus cache, where Docker defines cache as the inactive_file value (cgroup v2) or total_inactive_file (cgroup v1). We call that the working set. Samples come about every 2 to 3 seconds, so a spike shorter than that can be missed.
  • Median of 3 runs. Each run starts fresh and tears the stack down with its volumes. The table shows the median of the three idle figures and of the three run peaks, with the lowest and highest run in brackets.
  • Highest reading. Sizing does not use the median. It uses the highest reading in any run, counting the start-up peak as well, because a machine sized for the median fails on the day the app lands on the high side. For Uptime Kuma the median peak was 139 MiB and the highest reading 156 MiB; for Home Assistant, 368 MiB and 407 MiB.
  • The machine. A GitHub Actions ubuntu-latest runner with 2 vCPUs and 8 GB of RAM, a fresh one for each app. The CPU model is not under our control: the same runner type landed on AMD EPYC and Intel Xeon processors, so each row names its CPU, and the list under the table says which app ran on which. RAM carries over to other machines far better than speed. We never compare these numbers with other people’s.

The heaviest apps

Immich is the heavy one, and it is five rows because its memory depends on how you run it. The full stack (server, machine learning, Postgres and Valkey) idled at 1.3 GiB and peaked at 3.6 GiB while processing 20 uploaded photos. That is 8 GB by our rule.

Turning machine learning off is the biggest single lever. The stack without it idled at 1.1 GiB and peaked at 1.6 GiB in the same upload test, and the rule gives 4 GB instead of 8 GB. The saving is in the peak far more than at idle, because the machine-learning container is mostly idle between jobs and busy during them. Smart search and face detection do no work with it off. The details, and what each container does, are in Immich RAM requirements, measured, and the file is in the Immich Docker Compose guide.

Two other Immich rows are worth a sentence each. With every container capped at 4 GiB in total, the median peak was 3.0 GiB, and with machine learning off and a 2 GiB total cap, 1.5 GiB; the Immich page says what the caps did to the containers and how close they came. Limiting machine-learning jobs to one at a time gave a median peak of 3,685 MiB against 3,701 MiB for the default, and that page does not credit the setting with a saving.

Behind Immich, the next heavy ones are multi-container stacks with their own database. authentik idled at 1.0 GiB, which in our latest runs put it fifth by idle RAM, just above the Immich stack with machine learning off and a 2 GiB cap (1.0 GiB). That margin is about 4 MiB, far smaller than the spread between authentik’s own runs (840 to 1,071 MiB), so treat the two as level. Seafile idled at 616 MiB, against 56 MiB for Authelia. Authelia is also the clearest case of why you should read both columns: its peak was 607 MiB during 60 password logins, four at a time. The Authentik vs Authelia comparison separates what the data shows from what we infer about that gap, and the authentik Docker Compose guide has the file and the per-container figures.

The light ones

Five apps idled below 50 MiB: Pi-hole at 9 MiB, File Browser at 15 MiB, FileBrowser Quantum at 21 MiB, Vaultwarden with its Caddy proxy at 22 MiB and AdGuard Home at 26 MiB. They are the reason a small box can run a lot of services. The two DNS blockers are compared in AdGuard Home vs Pi-hole v6, the Pi-hole file and its setup notes are in the Pi-hole Docker Compose guide, the password server in the Vaultwarden Docker Compose guide and the two file managers in Self-hosted alternatives to Google Drive.

In between sit the apps people usually think of as small but that carry a real runtime: Uptime Kuma at 123 MiB idle, ownCloud Infinite Scale at 152 MiB, Jellyfin at 186 MiB, Nextcloud at 297 MiB and Home Assistant at 363 MiB. Home Assistant barely moves under its load (peak 368 MiB), while Jellyfin goes from 186 MiB to 406 MiB with one software transcode running. The guides: Uptime Kuma, Jellyfin, Nextcloud and Home Assistant. ownCloud Infinite Scale is covered in Nextcloud vs ownCloud vs Seafile.

How to add up a stack

Our sizing rule is written for one app. For several, we use the same arithmetic as the “Where to run it” section of our guides when it is given more than one app, and it is plain addition:

  1. Add the idle medians. That is the memory the stack holds all day.
  2. Add each app’s highest reading in any run (the figure behind its “RAM by our rule”).
  3. Multiply that second sum by 1.5 and round up to the next plan size, once, for the whole stack.

Here it is for five small services that people often run together:

Adding up the idle RAM and the highest readings of 5 apps, then applying the sizing rule
AppIdle RAM (MiB)Highest reading in any run (MiB)
Home Assistant363407
Vaultwarden with Caddy2269
Uptime Kuma123156
Pi-hole v6918
Jellyfin186406
Sum7041,056
Sum of the highest readings times 1.51,583 (1.5 GiB)
Plan size by the rule2 GB

And for a stack built around Immich with machine learning on:

Adding up the idle RAM and the highest readings of 4 apps, then applying the sizing rule
AppIdle RAM (MiB)Highest reading in any run (MiB)
Immich1,3213,802
Nextcloud297479
Home Assistant363407
Vaultwarden with Caddy2269
Sum2,0034,757
Sum of the highest readings times 1.57,135 (7.0 GiB)
Plan size by the rule8 GB

Be clear about what this is. We did not run these apps side by side, so the sums are arithmetic on separate measurements, not a measured total. Adding the highest readings assumes every app peaks at the same moment, which is the cautious assumption: it can overstate the need, and we did not test how often real peaks coincide. Some of our lab stacks bring their own database and cache (Immich has Postgres and Valkey, Nextcloud has Postgres and Redis). If you share one Postgres between apps, the sum is too high by an amount we have not measured. For whether a given box fits, start from the idle sum and see how far the second sum, times 1.5, is from your RAM.

What the numbers do not cover

  • The host. The totals are container memory. The operating system, the Docker daemon and anything else on the machine come on top. The 1.5 in our rule is our allowance for them and for spikes; we did not measure the host’s own use.
  • File cache. We count usage minus inactive file cache, so a media or file server with a large library will look fuller in free than in our table. The Linux kernel keeps file data it has read or written in the page cache and can reclaim cache pages when memory runs short, but reclaim takes work and written data must reach disk first. We did not test a library large enough to compete with the apps for RAM, so we cannot say how much the cache costs you in practice.
  • How many users, monitors or photos. Each load is one user and a fixed workload: 20 photos, 103 files, 30 monitors, 60 logins, one transcode. More of any of these raises the peak, and we have only measured the sizes named in the table. The Jellyfin guide adds a table of simultaneous transcodes on the 2-vCPU runner, which is the only multi-stream data we have.
  • Time. A run lasts minutes, not days. Memory growth over weeks, such as a database or cache that keeps growing, is not measured.
  • Other setups. ARM, GPU or hardware transcoding, other image variants, a reverse proxy in front (except Vaultwarden’s Caddy) and versions other than the one in the table.
  • Different CPUs. The rows ran on different CPU models, as listed under the table. We expect RAM to differ much less than speed between them, but we did not run one app on several models to check.

Download the data

Every row links to its raw result file (result JSON), which holds all three runs in full, the per-container figures, the CPU model of each run and the Docker and kernel versions. The guide linked in the same row has the verified Compose file and the full Test box with the start-up peak and the memory timeline. How the numbers are made is on the methodology page.

Update log

  • The authentik and Authelia rows follow their re-measurement together in one GitHub Actions job on one runner (Intel Xeon 6975P-C, Nov 8, 2026): authentik idle 1,067 MiB (was 836), peak 1,109 (was 1,014); Authelia idle 56 MiB (was 60), peak 607 (was 626). The table is built from the result files; authentik's idle now ranks fifth, above the Immich stack with machine learning off and a 2 GiB cap, so the heaviest-five check and one sentence about authentik changed.
  • Two rows follow the re-run of both DNS blockers in one GitHub Actions job on one runner (AMD EPYC 9V74), done in both run orders (Nov 5 and Nov 6, 2026; the table uses the Nov 6 job): AdGuard Home v0.107.79 (idle 26.4 MiB, was 39.3; peak 52.0, was 61.4) and Pi-hole 2026.10.1 (idle 9.3, was 9.4; peak 17.9, was 18.2). No tier changed: both rows are still 1 GB by the sizing rule. The AdGuard Home figure moved a lot with no change in the image (same digest), so it depends on the host. The prose is unchanged; the Pi-hole and AdGuard Home rows are now from the same job.
  • Five rows follow their re-tests of Nov 3, 2026: Jellyfin 12.4 (idle 186 MiB, was 191; peak 406, was 401), Uptime Kuma 2.5.9 (idle 123, was 129; peak 139, was 146), Vaultwarden 1.37.5 with Caddy (idle 22, was 25; peak 67, was 70) and Pi-hole 2026.10.1 (idle 9.4, was 9.3; peak 18.2, unchanged); the Vaultwarden backup lab is not in the table. No tier changed: every one of these rows is still 1 GB by the sizing rule. The CPU models changed for four of the five rows (Jellyfin and Uptime Kuma to an AMD EPYC 7763, Pi-hole to an Intel Xeon Platinum 8573C; Vaultwarden stayed on the EPYC 7763), and the runner image moved to Ubuntu 26.04.1 with Docker 29.4.2 (was Ubuntu 24.04.5 and Docker 28.0.4 for most rows), so we claim no cause for any change in the figures. The AdGuard Home row was not re-run, so the Pi-hole and AdGuard Home rows are now from different days. The prose is unchanged.
  • Linked the new authentik Docker Compose guide from the multi-container paragraph. No measurements changed.
  • Linked the new Pi-hole Docker Compose page from the light apps paragraph. No measurements or other prose changed.
  • The authentik row follows its re-test on 2026.8.5 (Oct 30; idle 836 MiB, was 942), which ran on a different CPU model from the first run. No prose changed.
  • First published. The table is built from the lab result files of the published apps, so it follows every re-test. It lists the apps with a service-type result on the site that day: Immich (five stacks), authentik, Seafile, Home Assistant, Nextcloud, Jellyfin, ownCloud Infinite Scale, Uptime Kuma, Authelia, AdGuard Home, Vaultwarden with Caddy, FileBrowser Quantum, File Browser and Pi-hole v6, from runs between Oct 8 and Oct 28, 2026. The backup tools (restic, Borg, Kopia) and the Vaultwarden backup lab are left out because their container figures describe a helper container or a backup job, not a service.