Catalogue vendor-minimum sweep (§45.6, ruled 2026-08-14)
Layer 6. Services. Services
One row per catalogue app that creates a container; the headless
packages are listed separately below, and between the two every package
under apps/ is accounted for. Cap is the primary (or tightest
vendor-relevant) service. “Vendor min” is a published figure for that
workload, not a host-size guess. “None published” means the vendor does
not document a RAM floor; the shipped cap stands.
Floors that this sweep raised are marked BREACH→fixed. Caps that already cleared a published floor are OK. Caps that sit on a published floor were raised (sonarqube precedent: a minimum is a floor to clear).
Guards: tests/unit/catalog/resource_floors_test.go (named + class
floors), tests/unit/catalog/host_sized_workers_test.go (the
by-construction nginx / JVM / Node mitigations), and
tests/unit/catalog/vendor_floor_table_test.go (this file’s COVERAGE;
an app added to apps/ with no row here turns it red by name).
| app | primary service | vendor min | shipped cap | verdict | source |
|---|---|---|---|---|---|
| activepieces | activepieces | 0.5 vCPU / 1 GB per concurrent flow (worker); app 1 vCPU / 1 GB per 10 workers | 4G | OK (single WORKER_AND_APP container carries both roles) | activepieces.com/docs/install/configure-operate/production-setup |
| static-site | nginx | none published (static files) | 256M | OK | none |
| stirling-pdf | stirling-pdf | none published; ~1 GB headroom for LibreOffice, ~2 GB for OCR on large scans | 4G (2G reservation) | OK (matches upstream compose) | upstream docker-compose; docs.stirlingpdf.com |
| onlyoffice | onlyoffice | none published for the fork; upstream ONLYOFFICE documents 2 GB min, 4 GB recommended | 4G (2G reservation) | OK (image bundles postgres+rabbitmq+converter pool in one container) | helpcenter.onlyoffice.com system requirements; Euro-Office README |
| affine | affine | 2 GB | 2G | OK (already floored) | docs.affine.pro self-host requirements |
| affine-rs | affine-rs | none for the Rust binary (Node affine is 2 GB) | 1G | OK; different binary; not the Node floor | none |
| agent-canvas | agent-canvas | none published | 4G | OK | none |
| blockscout | blockscout-backend | none published | 512M | OK | none |
| ciso-assistant | backend | none published | 1G | OK | none |
| ciso | backend | none published | 1G | OK | none |
| clickhouse | clickhouse | 4 GB | 4G | OK (already floored) | clickhouse.com/docs/operations/tips |
| crowdsec | crowdsec | none published | 512M | OK | none |
| defectdojo | defectdojo-uwsgi | none published | 1G | OK | none |
| defectdojo-classic | defectdojo-uwsgi | none published | 1G | OK | none |
| dependency-track | dtrack-apiserver | 2 GB / 2 CPU | 4G / 2.0 | OK (already floored) | docs.dependencytrack.org deploy-docker |
| dify | dify-plugin-daemon | none published | 2G | OK | none |
| docuseal | docuseal | none published | 1G | OK | none |
| drover-relay | drover-relay | none published (upstream happy-server documents no floor) | 1G | OK (one replica, Postgres on the estate’s own stack) | none |
| ethereum | geth | mainnet often 8 GB+; package posture unstated | 1G | LEFT; may be light/dev; not raised without package intent | geth.ethereum.org docs |
| fizzy | fizzy | none published (upstream’s Docker and Kamal guides size no floor) | 1G | OK | basecamp/fizzy docs/docker-deployment.md |
| fleet | fleet | 4 GB / instance | 4G / 1.0 | BREACH→fixed (was 512M / 0.5) | fleetdm.com reference-architectures |
| frappe | frappe-web | none published (bench-sized) | 4G | OK | none |
| ffx | frappe-web | none published (bench-sized) | 4G | OK | none |
| fusionpbx | fusionpbx | none published | 2G | OK | none |
| garage | garage-N | 1 GB RAM (stated requirement) | 1G | OK | garagehq.deuxfleurs.fr |
| gitea | gitea | ~512 MB–1 GB typical | 1.5G | OK | docs.gitea.com |
| gitlab | gitlab | 4 GB Omnibus min | 6G | OK; lowered from 16G on measurement (see note below) | docs.gitlab.com |
| gitlab-fizzy | gitlabfizzy | none published | 256M | OK | none |
| gitlab-huly | gitlabhuly | none published | 512M | OK | none |
| gitlab-runner | gitlab-runner | none published | 2G | OK | none |
| glitchtip | glitchtip-web | none published | 1G | OK | none |
| goalert | goalert | none published | 512M | OK | none |
| graphiti | neo4j | heap+pagecache+1 GB | 3G | OK (already floored) | neo4j operations-manual docker |
| greenbone | gvmd | none published (data images are large) | 4G | OK | none |
| hockeypuck | hockeypuck | none published | 512M | OK | none |
| hookshot | hookshot | none published | 512M | OK | none |
| huly | elastic / redpanda / nginx | ES 2G; Redpanda 2G; nginx idle floor 256M | 2G / 2G / 256M | OK (already floored) | elastic.co; redpanda sizing; measured |
| huly-matrix-notifier | notifier | none published | 512M | OK | none |
| hyperswitch | hyperswitch-server | none published | 512M | OK | none |
| identity-broker | broker | none published | 512M | OK | none |
| immich | immich-server / ML | 6 GB install / ML spike | 3G / 2G | OK (already floored) | docs.immich.app/install/requirements |
| ivatar | ivatar | none published | 768M | OK | none |
| jellyfin | jellyfin | ~2 GB typical | 2G | OK | jellyfin.org |
| journiv | journiv | none published | 1G | OK | none |
| keygen | keygen-web | none published | 1G | OK | none |
| kubo | kubo | none published | 1G | OK | none |
| kutt | kutt | none published | 512M | OK | none |
| lookout | lookout | none published (first-party One/node server; better-auth is stateless JWE) | 1G | OK | none |
| librechat | librechat | none published | 2G | OK | none |
| mailpit | none published | 256M | OK | none | |
| mailbox-reconciler | mailbox-reconciler | none published | 128M | OK | none |
| mailbox-reconciler-ffx | mailbox-reconciler | none published | 128M | OK | none |
| mailcow | stack (mysql+dovecot+clamd+…) | ~6 GB host typical | services sum ≫ 6G | OK | mailcow docs |
| mariadb | mariadb | none published | 1G | OK | none |
| matomo | matomo | none published | 512M | OK | none |
| matrix | synapse | 1 GB (large rooms) | 2G | sits-on-floor→cleared (was 1G) | matrix-org.github.io/synapse installation |
| matrix | jitsi-jvb | heap 768m inside 2G | 2G | OK (already floored + VIDEOBRIDGE_MAX_MEMORY) | jvb.sh default; issue #1575 |
| mautic | mautic | 2 GB host min | 2G | BREACH→fixed (was 512M) | mauteam.org hardware requirements |
| medusa | medusa | none published | 2G | OK | none |
| metabase | metabase | 1 GB typical | 4G | OK | www.metabase.com |
| midpoint | midpoint | ~2 GB typical | 4G | OK | docs.evolveum.com |
| midpoint-classic | midpoint | ~2 GB typical | 4G | OK | docs.evolveum.com |
| mongodb | mongodb | ~1 GB typical | 2G | OK | none |
| monica | monica | none published | 512M | OK | none |
| monica | monicanc | none published | 512M | OK | none |
| monica-nextcloud | monicanc | none published | 512M | OK | none |
| multiplatform | nginx | idle floor 256M | 256M | BREACH→fixed (was 128M) | measured; host_sized_workers |
| multiplatform-keycloak | keycloak | 1000 MB base for an idle instance, plus 500 MB per 100k active sessions | 1536M | OK (the heap follows the limit at MaxRAMPercentage=50; idles near 950M RSS on this image) | keycloak.org high-availability/concepts-memory-and-cpu-sizing; apps/integrated/multiplatform-keycloak/vars.yaml |
| mysql | mysql | none published | 1G | OK | none |
| netbird | netbird-management | none published | 512M | OK | none |
| nextcloud | nextcloud | 512 MB min / 2 GB rec | 2G | OK | docs.nextcloud.com |
| nginx | nginx | idle floor 256M | 256M | OK (already mitigated) | measured |
| ollama | ollama | model-dependent | 12G | OK | none |
| openarchiver | openarchiver | none published | 3G | OK; sized to the heap it authorises, not to a vendor figure: the container runs five node processes (frontend, backend, ingestion worker, indexing worker, sync scheduler) and hands the indexing worker INDEXING_WORKER_MAX_OLD_SPACE_MB=1536, so the cap is twice the heap | upstream .env.example; apps/standalone/openarchiver/vars.yaml |
| openarchiver | meilisearch | none published | 2G | OK | meilisearch.com docs |
| openarchiver | tika | none published | 1G | OK (JVM + tesseract-ocr on a scanned attachment) | tika.apache.org |
| openhands | openhands | 4 GB, published as a SYSTEM figure (“a system with a modern processor and a minimum of 4GB RAM is recommended to run OpenHands”) | 2G | OK; the cap is the ORCHESTRATOR only, and the vendor’s figure is not this cgroup’s floor: the app container binds /var/run/docker.sock and spawns each session’s SANDBOX_RUNTIME_CONTAINER_IMAGE on the HOST daemon, so the runtimes are siblings drawing host RAM outside this limit. Same reading as affine-rs (different subject, not the published floor) and mailcow (a host figure against a multi-service sum) | docs.openhands.dev/openhands/usage/run-openhands/local-setup |
| openreplay | chalice-openreplay | none published | 1G | OK | none |
| openreplay-ch | clickhouse | 4 GB | 4G | OK (already floored) | clickhouse.com/docs/operations/tips |
| paperless | paperless | ~1 GB typical | 2G | OK | docs.paperless-ngx.com |
| peertube | peertube | 1.5 GB | 2G | BREACH→fixed (was 1G) | joinpeertube.org/faq |
| penpot | penpot-frontend | idle floor (host-sized workers) | 1G | OK (already floored + autotune) | measured |
| plane | plane-api | none published | 2G | OK | none |
| pmm | pmm-server | ~2 GB typical | 4G | OK | docs.percona.com |
| postiz | postiz | none published | 1536M | OK | none |
| pulp | pulp-worker | none published | 3G | OK | none |
| redis | redis | none published | 512M | OK | none |
| s3-migrate | migrate | none published | 1G | OK | none |
| s3manager | s3manager | none published | 256M | OK | none |
| shlink | shlink | none published | 1G | OK | none |
| sonarqube | sonarqube | 4 GB Docker | 6G | OK (already floored) | docs.sonarsource.com linux pre-install |
| supabase | postgres / kong | none published as a stack floor | 1G | OK | none |
| superset | superset | none published | 2G | OK | none |
| uptime | uptime-kuma | none published | 512M | OK | none |
| uptime-kuma | uptime-kuma | none published | 512M | OK | none |
| vaultwarden | vaultwarden | none published (tiny) | 512M | OK | none |
| wazuh | wazuh-indexer | 1 GB heap → 2G cap | 2G | OK (heap raised 512m→1g; floor added) | documentation.wazuh.com indexer memory |
| whoami | whoami | none published | 256M | OK | none |
| windmill | windmill-server / worker | 2 GB stack; 2048M/worker | 2G / 2G | BREACH→fixed (was 512M / 512M) | windmill.dev self_host; official compose |
| wordpress | wordpress | ~512 MB typical | 512M | OK (sits on typical; no vendor floor) | none |
Note: gitlab, lowered 16G → 6G (2026-08-17)
Section titled “Note: gitlab, lowered 16G → 6G (2026-08-17)”This sweep only ever looked UPWARD: a cap that cleared its vendor minimum
was marked OK and left alone. gitlab’s 16G cleared the 4 GB Omnibus
minimum four times over and so was never examined, and the sweep’s own
framing (“shipped cap”) hid what the number really is. limits.memory in
an app compose is the quota, and the renderer derives both the
reservation and the limit from it: reservation = quota × deploy.reservation_ratio (1.0)
and limit = quota × deploy.limit_ratio (1.5). Placement admits a replica
when that reservation fits the node’s free capacity. A 16G quota is a 16 GiB reservation. On a 30.6 GiB
node it held 16 of 28.5 GiB of reservations, left 2.1 GiB schedulable, and
blocked a 4 GiB placement for a day while the node’s real usage was
~7.7 GiB.
Measured on the running package (puma worker_processes 1, prometheus
off, external postgres, gitaly in-container), from cgroup v2
memory.current / memory.peak: steady 3.18 GiB; memory.peak over
a 13.5 h window 3.30 GiB; with one full forked gitlab-rails runner
(the shape gitlab-backup create/restore takes) 3.98 GiB. No OOM
kill in 14 days of that node’s kernel log.
6G ⇒ reservation 6 GiB (1.51× the measured worst case) and limit 9 GiB (2.26× it, and above GitLab’s own 8 GB recommended figure). A cap far above a vendor minimum is not automatically safe; on this platform it is capacity withdrawn from every other deployment, so a large one now wants a measurement, not just a floor check.
Headless apps: no services, nothing to size
Section titled “Headless apps: no services, nothing to size”These twelve packages render services: {}. Each exists to declare a socket
onto an external or managed instance (RDS, a vendor SMTP relay,
someone else’s S3, a GitLab this estate does not administer), so it creates
no container, holds no cgroup, and has no cap a vendor minimum could floor.
They are absent from the table above by construction, not by oversight,
and they are named here so the next reader diffing the table against apps/
does not re-walk twelve packages to discover there was nothing in them.
external-gitlabexternal-imapexternal-llmexternal-mailcowexternal-mongodbexternal-mysqlexternal-oidcexternal-postgresexternal-redisexternal-s3external-smtps3
The exemption is declared rather than inferred on purpose: a package that
grows a real service while sitting on this list would otherwise keep a
silent pass from the sweep. vendor_floor_table_test.go fails by name in
that case, and in the reverse case of a row naming an app that is gone.
Class and by-construction floors
Section titled “Class and by-construction floors”Class floor (every *-pgclient): 512M / 0.5 cpu, measured; already enforced.
By-construction nginx mitigations (original six + later siblings):
penpot-frontend, huly/nginx, nginx/nginx, greenbone/nginx,
defectdojo/defectdojo-nginx, mailcow/nginx-mailcow, plus pulp-web,
openreplay frontend, shlink, dify, multiplatform. Guarded in
host_sized_workers_test.go.