Choose a deployment
Every outpost runs the same program and does the same work. What differs is who starts it, where its state lives, how you upgrade it, and how much of the host is yours to look after. Pick the deployment that matches how your team already runs hosts, not the one that looks most modern. This page gives you the comparison, a recommendation per kind of team, and the decisions that come after: how many outposts, how large, and what to do when one is down.
The four deployments
| Linux host with systemd | Docker | Docker Compose | AWS ECS | |
|---|---|---|---|---|
| What runs | The outpost binary as a systemd service, beside the host's Docker. | The outpost image as one container, with the Docker socket mounted. | The same container, described in two files. | A CloudFormation stack: one EC2 instance and one ECS task on it. |
| How it enrolls | outpost enroll once, then the service starts. | enroll in a one off container, then run. | On its first start, from the token in .env. | On its first start, from the token the stack puts in Secrets Manager. |
| State | /var/lib/outpost on the host. | /var/lib/outpost on the host, mounted at the same path. | Same as Docker. | /var/lib/outpost on the instance's disk. Lost if the instance is replaced. |
| Upgrade | Copy in the new binary, restart the service. | Remove the container, run the new tag. | Change the tag, pull, up -d. | Deploy the stack again with the new Image. |
| Stopping | Running scans get five minutes to finish. | Five minutes, with --stop-timeout 360. | Five minutes, from stop_grace_period. | 100 seconds: ECS stops a container after 120. |
| Logs | journald. | docker logs. | docker compose logs, rotated at 5 files of 20 MB. | CloudWatch Logs, kept 30 days. |
| Who patches the host | You. | You. | You. | You. The stack picks the current ECS optimised Amazon Linux 2023 image when it creates the instance and never replaces a running one. |
| HTTPS proxy | /etc/default/outpost, plus Docker's own proxy for image pulls. | Pass HTTPS_PROXY and NO_PROXY to the container, plus Docker's own proxy. | In .env, plus Docker's own proxy. | The HttpsProxy parameter sets it for the outpost, Docker and the ECS agent. |
| TLS inspecting proxy CA | --ca-file on enroll. | --ca-file on enroll, with the file mounted. | --ca-file on enroll, with the file mounted. | Not offered by the stack. |
| AWS identity for assumed roles | The host's credentials or instance profile. | The same, passed to the container. | The same, passed to the container. | The stack's task role, granted the roles you list in AssumeRoleArns. |
| Best fit | Platform teams with configuration management and a patching routine for Linux hosts. | A quick evaluation on a host you already run containers on. | A host you want described as files in version control. | AWS accounts where you would rather create and delete the outpost as a stack. |
All four need Docker Engine on the host, because the outpost runs every scan as containers on it. That is also why the AWS deployment uses ECS with EC2 capacity: Fargate gives a task no Docker to drive. See Outpost FAQ.
What we recommend
A platform team on ordinary Linux hosts: systemd. It fits the tools you already have: your configuration management installs a binary and a unit, journald collects the logs, and your patch routine covers the host. The outpost's process does not hold the Docker socket inside a container, which is one less thing to explain in a review. Choose Compose instead if your hosts are described as container files and nothing else.
An AWS first team: the ECS stack, sized on purpose. It creates and removes cleanly, keeps the enrollment token out of plain text, and gives the outpost an AWS identity of its own, which is what lets your AWS connections trust a role in your account instead of Dwaar's. Two things to decide before you deploy it: choose an instance type with the memory your scans need, because the default is too small for a repository scan with every category on, and plan how you patch the instance, because replacing it costs a new enrollment token. Behind a TLS inspecting proxy, run the outpost on an EC2 instance with systemd or Compose instead, until the stack can take your CA.
A single security engineer trying it out: Compose on a spare VM. Two files and one command, and docker compose down takes it away again. Give the VM 8 CPUs and 32 GiB if you want to see a full repository scan, and set its Memory overcommit to 1.1 in the Deploy outpost drawer, as Plan an outpost explains. A smaller VM still shows you discovery, connection tests and container image scans.
How many outposts
An outpost serves every connection bound to it, across every organization in your tenant. One is enough for many connections. Add another when one of these is true:
| Reason | What to do |
|---|---|
| The providers sit in networks that cannot reach each other | One outpost per network segment, on a host that reaches that segment's Git servers, registries or cloud APIs. |
| The scans queue behind each other for longer than you accept | Another outpost in the same segment, and split the connections between them. Or a larger host with a higher Scan concurrency. Each outpost's scans count only against its own Scan concurrency, never against Dwaar's limit for your tenant, so both shorten the queue. |
| A team must not run its scans on a host another team controls | An outpost per team. The host holds that team's credentials while its jobs run, so the host's owner is part of the trust. |
| Your data residency rules tie work to a region | An outpost in each region, binding each connection to the outpost in its region. |
More outposts are not free: each is a host to patch and watch. Start with one per network segment and add others when the queue tells you to.
Availability
There is no failover. A connection runs on exactly one outpost, the one under its Runs on, and Dwaar never sends its work anywhere else on its own. Two outposts in the same network do not share one connection's work and do not stand in for each other.
So plan for an outpost being down the way you plan for any single host:
| When the outpost is down | What happens | What you do |
|---|---|---|
| For a few minutes, such as a restart or an upgrade | Scans wait and run when it is back. Discovery and tests in that window fail and run again at the next discovery. | Nothing. |
| For hours, such as a failed host | Scans that already started on it wait for it, then fail. Discovery and tests fail. | Move the connections to another outpost under Where it runs. New work goes there at once. |
| For good | As above. | Revoke it, then move its connections. See Revoking and replacing. |
If a few hours of missed scans matters to you, keep a second outpost enrolled and online in the same network, with nothing bound to it. Moving connections to it is a few clicks per connection and needs no new token. Size it like the first, or it will refuse the scans you move to it.
Which outpost a connection uses
You choose it per connection, under Where it runs, when you connect the provider or later on the connection's Connection tab. The picker offers Dwaar and your outposts: in the connect form those that are online or just enrolled, on the Connection tab every one that is not revoked. Choose the outpost that:
- can reach the provider: the same network as the Git server or registry, or an AWS identity the connection's role trusts,
- is large enough for the scans this connection will send, and
- belongs to the team whose credentials it is.
Test connection runs on the outpost you chose, so a passing test proves the first point. The Outposts tab lists each outpost's Bound connections, so you can see the load you are adding to. See Bind connections.
Try this next
Write down, for each provider you want behind an outpost, the network segment it lives in and the team that owns its credentials. The number of distinct pairs is the number of outposts you need. If it is more than three, start with the pair that blocks the most connections today.