Skip to content

Yet another process supervisor but with JSON logs and Vault integration

Notifications You must be signed in to change notification settings

mcrute/simplevisor

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Yet Another Process Supervisor

Simplevisor is yet another process supervisor but with a very narrow scope and a specific set of requirements that caused its creation. If you want a full-featured process supervisor you probably want something other than this, but if you share these requirements then you may have found what you need.

The requirements for Simplevisor are:

  • Be a suitable PID 1 for containers (handle signals and shutdown)
  • Allow running processes as non-root users
  • Allow running jobs that initialize the environment
  • Keep jobs alive
  • Support consolidated JSON logging for all sub-processes
  • Allow re-mapping termination signals
  • Support integration with Vault for Vault-unaware applications (bonus points for 12-Factor patterns)

And this is what Simplevisor does. Nothing more, nothing less.

Still To Do

  • Process restarts with backoff do not work
  • Signals need better testing
  • Vault token passing does not work
  • Support processes that already log in JSON format

Usage

Simplevisor tries to be very simple to use. If the config file simplevisor.json is in the same directory as the binary then running ./simplevisor as root with some Vault environment variables is sufficient to start all of the processes and keep them running.

Vault integration is enabled by default, but can be disabled by passing --no-vault. If Vault integration is enabled then the following variables must be present in the process environment. VAULT_ADDR, which must contain a URL pointing to Vault. One of either VAULT_TOKEN, containing a token to use when authenticating Vault calls, or both of VAULT_ROLE_ID and VAULT_SECRET_ID when using AppRole authentication.

In addition to disabling Vault integration --config can be passed to provide a non-standard location for the config file.

Configuration

The configuration file format is JSON and is well documented (for full details see: supervisor/model.go). Here is a worked example.

Environment

By default environment variables are only passed through to subprocesses if they are on the pass list. This prevents leaking secret environment variables meant for Simplevisor itself through to the managed processes. This can be overridden by setting pass-all to true. All variables to be passed through to the managed process must be in pass, variables in the Vault specific lists do not imply their existence in pass.

Vault

IMPORTANT: The Vault integration requires periodic login tokens. Not using periodic tokens will cause Simplevisor to eventually fail to renew the credential lease and terminate all managed processes.

Vault integration works by looking for variables in the Simplevisor environment named in the vault-replace and vault-template keys. If keys are found that match vault-replace they will be parsed, looked up in Vault, and the returned value will be injected into the environment of managed processes. This will happen once at Simplevisor startup time and each managed process will see the same secrets.

Replacement variables are colon (:) separated lists of three arguments: path, field, and type. The path is the path within the mount-point to the secret. For secret type credential the field refers to the key in the returned JSON document, only single-level JSON documents containing string keys and values are supported. Other types have constraints listed below.

Type is one of:

  • db, which refers to database credentials (mounted at database/ in Vault). Valid field names are Username and Password, nothing else.
  • secret, which refers to JSON formatted KV secrets (mounted at kv/ in Vault).
  • aws-user, which refers to an AWS IAM user credential (mounted at aws/ in Vault). Valid field names are KeyId and SecretKey, nothing else.

Credentials are cached upon first fetch and subsequent references to them will used the cached value. This presents a consistent view of the secrets to the process regardless of how many times the secret is referred to in the environment.

Simplevisor will manage renewing the login token and any credential leases that are acquired as part of the environment expansion described here.

For example, given the configuration below and the following environment state when Simplevisor is launched:

MONGO_USER="db:prod-db:Username"
MONGO_PASSWORD="db:prod-db:Password"
DJANGO_SECRET="secret:my/app/django-secret:Secret"

Assuming that kv/v1/my/app/django-secret contains:

{ "Secret": "some-secret" }

Managed processes would observe the following in their environment:

MONGO_USER="some-username-for-prod-db"
MONGO_PASSWORD="password-for-the-above-username"
DJANGO_SECRET="some-secret"

Vault Templates

Templates are designed to allow creation of more complex secrets based on other secrets, such as JDBC connection strings. The vault-template list contains variables that, if found in the Simplevisor environment, will be parsed as Go text/template templates and rendered with a context of the resolved variables in vault-replace. Note that the templates will not have access to the full environment.

For example the following template, evaluated with the variables from the prior section:

MONGO_URL="mongodb://{{ .MONGO_USER }}:{{ .MONGO_PASSWORD }}@my-mongo-host.prod:27017/prod-database?authSource=admin"

This would result in managed processes observing:

MONGO_URL="mongodb://some-username-for-prod-db:[email protected]:27017/prod-database?authSource=admin"

Vault Token

The supervisor Vault token can be injected to managed processes as VAULT_TOKEN. This implies that VAULT_ADDR will also be injected.

Job Configuration

There are two types of jobs init jobs and main jobs. They differ only in when and how they are run and if they are restarted on failure. All jobs receive an identical environment, per the preparation noted above.

init jobs are run serially before the main jobs are started. They are expected to exit with a zero status code. The failure of an init job will result in the termination of all jobs and the supervisor exiting with error.

main jobs are configured identically to init jobs but are started in parallel after the init jobs complete and are restarted if they fail. Failure of a main job does not terminate the supervisor. Failing main jobs are restarted with exponential backoff.

Each job will be spawned as the leader of its own session and will run as the configured user and group. If user and group are not specified then root:root is assumed. If user is specified but group is not then root is assumed. The run-as key takes the form user:group.

When the process supervisor is shutting down it will send a TERM signal to all managed processes. This can be configured with the kill-signal flag which should be the signal name without the SIG prefix.

Full Config Example

{
    "env": {
        "pass": [
            "PATH",
            "HOME",
            "PWD",
            "MONGO_URL"
        ],
        "vault-token": false,
        "pass-all": false,
        "vault-replace": [
            "DB_USERNAME",
            "DB_PASSWORD",
            "DJANGO_SECRET_KEY"
        ]
        "vault-template": [
            "MONGO_URL"
        ]
    },
    "jobs": {
        "init": [
            {
                "cmd": ["/setup-env.sh"],
                "run-as": "netbox"
            }
        ],
        "main": [
            {
                "name": "queue-worker",
                "cmd": ["/usr/bin/python3", "/opt/netbox/netbox/manage.py", "rqworker"],
                "run-as": "netbox"
            },
            {
                "cmd": ["/usr/sbin/uwsgi", "--ini", "/etc/uwsgi/netbox.ini"],
                "kill-signal": "INT",
                "run-as": "root"
            }
        ]
    }
}

Logging

When a managed process writes to either stdout or stderr those log messages will be captured, attributed to the process, and logged to the stdout of Simplevisor as JSON. The JSON format contains the fields:

  • process: the name of the process, either as configured in the config file or the basename of the first argument in the command. If the command name is not descriptive or ambiguous, set the name in the job configuration. This is not mandatory but will make reading logs easier.
  • time: the Unix timestamp of the log entry in integral format
  • stream: an integer indicating the stream the process wrote to 0 for stdout, 1 for stderr.
  • message: one line of the message written by the process

Writing multiple lines will result in multiple log messages (as can be the case for stack traces). Lines are always newline terminated.

Example:

{"process":"bash","time":1670346907,"stream":0,"message":"..."}
{"process":"internal","time":1670349781,"stream":0,"message":"..."}

But Why?

This all seems pretty complex and a lot of moving pieces, and in a sense it is. This is also a major simplification of what existed before. Prior to Simplevisor we used an amalgamation of dumb-init, runit, and su-exec with a liberal amount of shell scripting and, in more complex cases, custom process wrappers to properly integrate processes with our infrastructure systems. In a few cases third-party applications required us to carry local patches anywhere from 50-1000 lines to support this which added a lot of burden in upgrading these applications and in some cases were very fragile.

Simplevisor has fixed all of this. It's eliminated all first-party patches, most of the shell scripts, and all of those extra tools in favor of one small binary (~15Mb on disk and ~13Mb RAM at steady-state) and a config file added to our application distributions.

Can something else do this? Probably. There are a lot of process supervisors out there and this one certainly isn't the best but we think the consolidated logging as structured JSON and Vault integration really set this one apart.

Contributing

We would very much appreciate bug fixes and other contributions provided they fit of the stated goals of the project.

If you find anything here useful and would like to submit patches please email the patch (in git format-patch format) or a repository location and branch name that the maintainers can pull and merge. We reserve the right to request changes to patches or reject them outright but are most likely to willing and thankfully merge them if they fit into the general theme here.

Contributors

About

Yet another process supervisor but with JSON logs and Vault integration

Topics

Resources

Stars

Watchers

Forks

Packages

No packages published