xqute

A job management system for python

Features

• Written in async
• Plugin system
• Scheduler adaptor
• Job retrying/pipeline halting when failed
• Support cloud working directory
• Support Google Batch Jobs scheduler

Installation

pip install xqute

A toy example

import asyncio
from xqute import Xqute

async def main():
    # 3 jobs allowed to run at the same time
    xqute = Xqute(forks=3)
    for _ in range(10):
        await xqute.put('sleep 1')
    await xqute.run_until_complete()

if __name__ == '__main__':
    asyncio.run(main())

API

https://pwwang.github.io/xqute/

Usage

Xqute object

An xqute is initialized by:

xqute = Xqute(...)

Available arguments are:

• scheduler: The scheduler class or name
• plugins: The plugins to enable/disable for this session
• workdir: The job meta directory (Default: ./.xqute/)
• forks: The number of jobs allowed to run at the same time
• error_strategy: The strategy when there is error happened
• num_retries: Max number of retries when job_error_strategy is retry
• submission_batch: The number of consumers to submit jobs
• scheduler_opts: Additional keyword arguments for scheduler
• jobname_prefix: The prefix of the job name

Note that the producer must be initialized in an event loop.

To push a job into the queue:

await xqute.put(['echo', 1])

Using SGE scheduler

xqute = Xqute(
    'sge',
    forks=100,
    scheduler_opts=dict(
        qsub='path to qsub',
        qdel='path to qdel',
        qstat='path to qstat',
        q='1-day',  # or qsub_q='1-day'
    )
    ...
)

Keyword-arguments with names starting with sge_ will be interpreted as qsub options. list or tuple option values will be expanded. For example: l=['h_vmem=2G', 'gpu=1'] will be expanded in wrapped script like this:

# ...

#$ -l h_vmem=2G
#$ -l gpu=1

# ...

Using Slurm scheduler

xqute = Xqute(
    'slurm',
    forks=100,
    scheduler_opts = {
        "sbatch": 'path to sbatch',
        "scancel": 'path to scancel',
        "squeue": 'path to squeue',
        "partition": '1-day',  # or partition='1-day'
        "time": '01:00:00',
        ...
    },
)

Using ssh scheduler

xqute = Xqute(
    'ssh',
    forks=100,
    scheduler_opts={
        "ssh": 'path to ssh',
        "servers": {
            "server1": {
                "user": ...,
                "port": 22,
                "keyfile": ...,
                # How long to keep the ssh connection alive
                "ctrl_persist": 600,
                # Where to store the control socket
                "ctrl_dir": "/tmp",
            },
            ...
        }
    },
    ...
)

SSH servers must share the same filesystem and using keyfile authentication.

Plugins

To write a plugin for xqute, you will need to implement the following hooks:

• def on_init(scheduler): Right after scheduler object is initialized
• def on_shutdown(scheduler, sig): When scheduler is shutting down
• async def on_job_init(scheduler, job): When the job is initialized
• async def on_job_queued(scheduler, job): When the job is queued
• async def on_job_submitted(scheduler, job): When the job is submitted
• async def on_job_started(scheduler, job): When the job is started (when status changed to running)
• async def on_job_polling(scheduler, job): When job status is being polled
• async def on_job_killing(scheduler, job): When the job is being killed
• async def on_job_killed(scheduler, job): When the job is killed
• async def on_job_failed(scheduler, job): When the job is failed
• async def on_job_succeeded(scheduler, job): When the job is succeeded
• def on_jobcmd_init(scheduler, job) -> str: When the job command wrapper script is initialized before the prescript is run. This will replace the placeholder {jobcmd_init} in the wrapper script.
• def on_jobcmd_prep(scheduler, job) -> str: When the job command is right about to run in the wrapper script. This will replace the placeholder {jobcmd_prep} in the wrapper script.
• def on_jobcmd_end(scheduler, job) -> str: When the job command wrapper script is about to end and after the postscript is run. This will replace the placeholder {jobcmd_end} in the wrapper script.

Note that all hooks are corotines except on_init, on_shutdown and on_jobcmd_*, that means you should also implement them as corotines (sync implementations are allowed but will be warned).

You may also check where the hooks are called in the following diagram:

To implement a hook, you have to fetch the plugin manager:

from simplug import Simplug
pm = Simplug('xqute')

# or
from xqute import simplug as pm

and then use the decorator pm.impl:

@pm.impl
def on_init(scheduler):
    ...

Implementing a scheduler

Currently there are a few builtin schedulers: local, slurm, gbatch and sge.

One can implement a scheduler by subclassing the Scheduler abstract class. There are three abstract methods that have to be implemented in the subclass:

from xqute import Scheduer


class MyScheduler(Scheduler):
    name = 'mysched'

    async def submit_job(self, job):
        """How to submit a job, return a unique id in the scheduler system
        (the pid for local scheduler for example)
        """

    async def kill_job(self, job):
        """How to kill a job"""

    async def job_is_running(self, job):
        """Check if a job is running"""