Operations
Limits of this version
This page lists what Podos does not do, as of the code on the main branch
(milestones M0 to M5). There are still no release binaries: you build
podos from source, or, Available once a release exists, use the
install script. The same list is printed by:
podos version --limits
If you upgrade a server that was set up with an M2 build, see Upgrading from an M2 server.
Runtime
| Limit | Planned fix |
|---|---|
Available Podos needs Linux 5.6 or newer, systemd 245 or newer, cgroup v2, crun, conmon and nft. |
|
Available A service without resources.memory gets 512 MiB, with no swap beyond it. A process that goes over it is killed by the kernel and shown as oom-killed. Set resources.memory for databases, JVMs and build steps. |
|
| Available Containers can use TCP and UDP over IPv4 (and ping), Unix sockets, and the routing netlink to list interfaces. Packet capture, DHCP clients, raw sockets, other netlink protocols, SCTP, AF_ALG, vsock and Bluetooth are not available. IPv6 inside application networks is off. | |
| Available The seccomp profile covers only the 64-bit architecture of the server. 32-bit x86 images do not run on x86_64 servers. On arm64 servers, 32-bit ARM images are allowed. | |
| Available Volumes belong to one environment and are bind-mounted from the server's data directory. Removing a service or environment never deletes volume data unless you ask for it. | volume backup Planned · 0.2 |
Available Files are at most 1 MiB each, 64 per environment and 16 per service, mounted read-only at one path each. No directory of files, no file from the repository, no file inside a volume or over /etc/hosts, /etc/hostname or /etc/resolv.conf. A change takes effect at the next deploy, not in running containers. See Files. |
|
| The disk space a container writes is not limited per container. Available Podos refuses to create containers when the data disk has less than 1 GiB or 5% of inodes free. | |
| Available Logs keep roughly the last 10 MiB per container. Each container has one log file of at most 10 MiB. At the limit the file starts empty and older lines are lost. The size is not configurable. | persistent logs and metrics Planned · 0.2 |
| Containers run as host root, without user namespaces, with a reduced capability set, a seccomp profile and no-new-privileges. | user namespaces after 0.1 |
| Services must keep running or be restarted by their restart policy; run-to-completion jobs are not supported. | not yet scheduled |
Available Health checks are http or tcp. Exec (command) checks are refused. |
Planned |
Available build needs BuildKit and crun 1.15 or newer on the server; the install script provides both. Without them a release with build is refused. |
|
| Images are pulled anonymously; no private registries. | not yet scheduled |
| No namespace quotas and no registry allow-list. | Planned |
Node protection
Available At every start the daemon reads the server's memory, CPU count and disk, because servers get resized. Then:
- All containers together may use at most the server's memory minus a system
reserve. The default reserve is 10% of the memory, at least 256 MiB and at
most 2 GiB; change it with
podos daemon --system-reserve-memory. - All containers together may use at most 80% of the kernel's process limit.
- A service whose
resources.memoryis larger than what the server can give to containers is refused at deploy. - The disk guard refuses a deploy or an image pull with "not enough free disk
space on the server" when the data directory's file system has less free
space than the threshold. The default is the larger of 1 GiB and 5% of the
file system, at most 10 GiB; change it with
podos daemon --min-free-disk. The exact numbers are in the daemon log, not in the error. Running containers are never stopped because of it. - Secret files have their own memory budget, counted inside the system reserve. See Files.
Files
Available What files in containers limit:
| What | Limit |
|---|---|
| One file, also a template once rendered | 1 MiB |
| Files in one environment | 64 |
| Files mounted by one service | 16 |
| Different values in one template | 256 |
| Stored content of a namespace, including what old releases still pin | 256 MiB |
file set and file unset per token |
60 per minute |
A template is rendered when a container is created, with the values the environment holds then, so a restart or a rollback uses today's values.
Files memory
Plain files are written to the data disk with the container's other state. Secret files, and templates that hold a secret once rendered, are written only to a memory file system of the daemon, never to the data disk.
- The memory file system has a fixed size for the whole server, shared by all
namespaces and environments. The default is 2% of the server's memory, at
least 16 MiB and at most 256 MiB. Change it with
podos daemon --files-memory. - It is counted inside the system reserve (see Node protection),
not taken from the containers' memory. It may be at most half of the
reserve; the daemon refuses to start with a larger value and says to lower
--files-memoryor raise--system-reserve-memory. - Every replica has its own copy of its secret files.
- One release may use at most a quarter of the budget. While a release rolls
out, old and new replicas both hold copies, so the release is counted
twice. A release that needs more is refused at deploy, with a message that
names the services; use fewer or smaller secret files or fewer replicas, or
raise
--files-memory. - On Linux 6.4 or newer the memory file system is not swapped out. On an older kernel its pages can reach swap unless swap is encrypted or off, and the daemon logs a warning.
- Plain files are copied for every replica too: the copies of one release may take at most 1 GiB of disk, and with the active releases of the namespace's other environments at most 2 GiB.
Builds on the server
Available What build does and does not do. See
Build from source for the details.
- Not supported: build arguments, secrets in a build, build targets, other
platforms than the server's own, private base images (no registry
credentials reach a build), and builders other than a
Dockerfile. - The Dockerfile must be named
Dockerfileand sit at the root of the build directory. - Defaults, changed with
podos daemonflags: one build takes at most 15 minutes and a source at most 256 MiB; the build cache is capped at 10 GiB; all builds together get half of the server's memory (at least 512 MiB, at most 8 GiB); as many builds run at once as the server has CPUs, at most 4. A build waits at most 10 minutes for its turn, and one token may start 10 builds per minute. - Builds of different namespaces never run at the same time. One namespace may have at most the number of builds that run at once plus two in progress.
- A build is cancelled when the free disk space falls below the server's minimum.
- Known gaps: images built on the server are never cleaned up yet, and a failed first build leaves an empty namespace.
Access
| Limit | Planned fix |
|---|---|
| The CLI works only on the server, as root, over the Unix socket. | Available HTTPS API, tokens, podos login, contexts |
| One server only. | Planned · 0.4 several nodes |
Available The web UI shows the server, services, domains, variables and secrets, and changes variables and secrets. Deploys, releases, files, tokens, server settings, logs and running commands are CLI only. A binary built without make ui has no UI. |
logs and metrics screens Planned · 0.2 |
| Available The install script and its systemd unit are built, but there is no published release to install from yet. Install from a source build today. | the first release |
Remote access
Available These limits apply to the remote access over HTTPS:
| Limit | Planned fix |
|---|---|
| No CA rotation: the internal CA cannot be replaced yet. | not yet scheduled |
| HTTP proxies are not supported for server connections; proxy variables are ignored. | not yet scheduled |
A server upgraded from M2 has no admin token. Create one as root on the server with podos token create --name NAME --kind admin. |
|
| A deploy token can read its environment's secrets through the containers it deploys. | |
| Several servers are separate: no hub, no combined view. | Planned · 0.2 podos ui |
Networking, restarts and rollouts
Available These are the facts behind the limits above.
- Each namespace and environment pair has its own private network, and services reach each other by name. Other namespaces and environments are never reachable unless shared explicitly (Planned · 0.3).
- Containers reach the internet and, by default, the server's private
networks. A stricter mode is chosen at setup (
podos setup --strict-egress). The cloud metadata addresses are always blocked, also in strict mode with an allow-list (see Security model). IPv6 inside application networks is off. - A domain belongs to one environment; the first claim wins across the server.
- Two environments cannot publish the same host port, and a port the server
already listens on (such as 22) cannot be published; the message
is the same for every reason. A service with
publishneeds one replica, and the port is not reachable from the server itself. - A container that stops by itself is restarted according to
restart, with a delay that grows from 1 second to 5 minutes. After a reboot, all containers of the active releases come back; a restart of the daemon does not restart containers. - A rolling update starts a new replica before it stops an old one. It fails and rolls back when a service has zero ready replicas for 15 seconds within 60 seconds of a deploy; the previous release is restored.
tls = "auto"needspodos setup --acme-emailand has only been tested against a test ACME server, not the real Let's Encrypt service. A wildcard domain cannot use it.podos setupchecksnftand the kernel version. It does not check the systemd version.
Upgrading from an M2 server
An M2 build ran containers with only a loopback network, no restart, did not
act on port, domains, tls, publish and health, and refused
volumes. All of that is now in place. A server set up with M2 needs
nftables and has no admin token; see
Upgrading an M2 server. Available Containers
started by a version before M3 keep running after an upgrade but without
networking until the application is deployed again.
Validation limits
| What | Limit |
|---|---|
| Manifest file | 256 KiB |
| Services per manifest | 200 |
| Environments per manifest | 20 |
| Replicas per service | 1000 |
env entries per service |
1000 |
env value |
32 KiB |
podos logs --tail |
10000 lines per replica |
| Concurrent log streams | 64 per daemon, 8 per environment |