Basic operation

A queue is specified by a storage account and a queue name:

export JOBQ="mystorageaccount/test-queue"

You should pick a name that avoids conflicts with other people. All the tasks on a queue are expected to run in the same “environment” and on the same “hardware”. If you want to send some tasks to GPU workers and other tasks to CPU workers, use separate queue names.

Here, tasks are represented as bash commands.

# send a single task
ai4s-jobq $JOBQ push -c "echo hello"

# execute a single task
ai4s-jobq $JOBQ pull

# pull and execute tasks in a loop
ai4s-jobq $JOBQ worker

Pull the first-pushed job from the queue and execute it. - If there are no jobs left, ai4s-jobq exits with code 0. If the worker is running on Azure ML, the job would succeed. - If the job fails, it is put back in the queue, until the max number of retries is exceeded.

If a worker dies while processing a task, the task will reappear at the end of the queue after --visibility-timeout. (This functionality is provided by Azure queues.)

When pushing commands, you can also: - specify environment variables, eg. -e AMLT_OUTPUT_DIR=/mnt/default/some/dir - enable background directory syncing like in amulet --bg-dirsync-to /mnt/default/some/dir, and let your job write to $AMLT_DIRSYNC_DIR.

Queueing many jobs from CLI

When running ai4s-jobq push without the -c option, it will read commands from standard input until EOF. Each line corresponds to one task. You can either specify shell commands, or kwargs as json. kwargs should start with an opening brace. For example, the following two tasks do the same thing:

cat tasks.txt
{"cmd": "echo Hello $NAME", env={"NAME": "John Doe"}}
echo Hello John Doe

# enqueue both of them
cat tasks.txt | ai4s-jobq $JOBQ push