devcluster

devcluster

A developer tool for running the Determined cluster.

Configuring and Running

You can run the code from the git root by running python -m devcluster, or install it into your python environment with pip install -e . and then just call devcluster by itself.

The first time you run it, it will give you the option to automatically install the default configuration file (see devcluster/example.yaml).

By default, devcluster needs to run from the root of a determined source tree. You can do this in one of a few ways:

• always run devcluster from the root of your determined source tree
• set cwd: /path/to/determined to run devcluster from anywhere
• run devcluster with the -C /path/to/determined option

You can either specify a config file via the --config or -c option, or devcluster will try to read the default config file at ~/.devcluster.yaml.

How devcluster Works

devcluster runs a cluster with a linear dependency graph of stages. With the default config (devcluster/example.yaml), those stages are:

• DEAD
• DB
• MASTER
• AGENT

Because of the linear dependency graph, the DB stage can run by itself, the MASTER stage is only started after DB is already up, and the AGENT is only started after the MASTER is up. The DEAD stage indicates that nothing is running. The non-DEAD stages correspond to the stages setting in your config file.

The status bar has two rows. The first row is the state row. It shows all of the stages in your dependency graph. All active stages are highlighted orange. There is also a 'target state', which is indicated by a less-than character (<). You set the target state by pressing the key corresponding to a particular stage.

When the target state is modified, devcluster will walk up or down the stage list, starting or killing processes as needed to reach the target stage. In the default configuration, the MASTER and AGENT stages are rebuilt each time they are started.

A common pattern during development might be to:

• make a change to the master code
• press 1 to set the target state to DB (killing the MASTER and the AGENT)
• press 3 to set the target state back to AGENT (rebuilding/restarting the MASTER and AGENT on the way).

The second row in the status bar is the logging row. Every stage collects a separate stream of logs, and each stream can be toggled by pressing the corresponding keys.

Keybindings

• 1-3 set target states for the cluster (as shown in the status bar)
• !-# toggle logs for the corresponding stage (as shown in the status bar)
• q or ctrl+c once to quit, or twice to force-quit (which may leave dangling processes or docker containers laying around)
• u/d will scroll up/down, and U/D will scroll up/down faster Scrolling support is only partially line-aware; it will scroll by literal log chunks, which may each be more or less than a line
• x will reset scrolling to continue following logs
• Spacebar will emit a visual separator with a timestamp to the console stream.
• With the default configuration, the keys p, w, and c will trigger building the python harness, webui, and docs, respectively.

Server/Client Mode

You can run devcluster as a headless server with various configurations for listening for clients:

# listen on 0.0.0.0:1234
devcluster --quiet --listen 1234

# listen on localhost:1234
devcluster -q -l 127.0.0.1:1234

# listen on a unix socket
devcluster -q -l /path/to/socket

# listen on the default unix socket (/tmp/devcluster/sock)
devcluster -q

# configure multiple listeners
devcluster -q -l 1234 /path/to/socket

Sending a SIGINT (via ctrl+c) or a SIGTERM (via kill) to the headless server will close it.

You can then connect the UI to a headless server:

devcluster 1234
devcluster host:1234
devcluster /path/to/socket

In client mode, pressing q or ctrl+c will only close the UI; it will not affect the server.

Oneshot Mode

You can run devcluster in a special oneshot mode (-1/--oneshot), suitable for automated usage. In automated mode, the stdout from devcluster will be suitable for redirecting to a file, and the stderr will emit the line:

devcluster is up

to indicate when the cluster is ready for use.

Philosophy

devcluster tries as much as possible to be a standalone tool for running arbitrary versions of the Determined cluster. It is tested as far back as v0.12.3.

Some defaults are kept up-to-date with tip-of-main to simplify the config file for the most-common use case, but in general preference was given to being configurable and explicit.

TODO

• config file reloading