Synnefo » snf-ganeti

============

Chained jobs

============

.. contents:: :depth: 4

This is a design document about the innards of Ganeti's job processing.

Readers are advised to study previous design documents on the topic:

- :ref:`Original job queue <jqueue-original-design>`

- :ref:`Job priorities <jqueue-job-priority-design>`

- :doc:`LU-generated jobs <design-lu-generated-jobs>`

Current state and shortcomings

==============================

Ever since the introduction of the job queue with Ganeti 2.0 there have

been situations where we wanted to run several jobs in a specific order.

Due to the job queue's current design, such a guarantee can not be

given. Jobs are run according to their priority, their ability to

acquire all necessary locks and other factors.

One way to work around this limitation is to do some kind of job

grouping in the client code. Once all jobs of a group have finished, the

next group is submitted and waited for. There are different kinds of

clients for Ganeti, some of which don't share code (e.g. Python clients

vs. htools). This design proposes a solution which would be implemented

as part of the job queue in the master daemon.

Proposed changes

================

With the implementation of :ref:`job priorities

<jqueue-job-priority-design>` the processing code was re-architectured

and became a lot more versatile. It now returns jobs to the queue in

case the locks for an opcode can't be acquired, allowing other

jobs/opcodes to be run in the meantime.

The proposal is to add a new, optional property to opcodes to define

dependencies on other jobs. Job X could define opcodes with a dependency

on the success of job Y and would only be run once job Y is finished. If

there's a dependency on success and job Y failed, job X would fail as

well. Since such dependencies would use job IDs, the jobs still need to

be submitted in the right order.

.. pyassert::

   # Update description below if finalized job status change

   constants.JOBS_FINALIZED == frozenset([

     constants.JOB_STATUS_CANCELED,

     constants.JOB_STATUS_SUCCESS,

     constants.JOB_STATUS_ERROR,

     ])

The new attribute's value would be a list of two-valued tuples. Each

tuple contains a job ID and a list of requested status for the job

depended upon. Only final status are accepted

(:pyeval:`utils.CommaJoin(constants.JOBS_FINALIZED)`). An empty list is

equivalent to specifying all final status (except

:pyeval:`constants.JOB_STATUS_CANCELED`, which is treated specially).

An opcode runs only once all its dependency requirements have been

fulfilled.

Any job referring to a cancelled job is also cancelled unless it

explicitly lists :pyeval:`constants.JOB_STATUS_CANCELED` as a requested

status.

In case a referenced job can not be found in the normal queue or the

archive, referring jobs fail as the status of the referenced job can't

be determined.

With this change, clients can submit all wanted jobs in the right order

and proceed to wait for changes on all these jobs (see

``cli.JobExecutor``). The master daemon will take care of executing them

in the right order, while still presenting the client with a simple

interface.

Clients using the ``SubmitManyJobs`` interface can use relative job IDs

(negative integers) to refer to jobs in the same submission.

.. highlight:: javascript

Example data structures::

  # First job

  {

    "job_id": "6151",

    "ops": [

      { "OP_ID": "OP_INSTANCE_REPLACE_DISKS", ..., },

      { "OP_ID": "OP_INSTANCE_FAILOVER", ..., },

      ],

  }

  # Second job, runs in parallel with first job

  {

    "job_id": "7687",

    "ops": [

      { "OP_ID": "OP_INSTANCE_MIGRATE", ..., },

      ],

  }

  # Third job, depending on success of previous jobs

  {

    "job_id": "9218",

    "ops": [

      { "OP_ID": "OP_NODE_SET_PARAMS",

        "depend": [

          [6151, ["success"]],

          [7687, ["success"]],

          ],

        "offline": True, },

      ],

  }

Implementation details

----------------------

Status while waiting for dependencies

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Jobs waiting for dependencies are certainly not in the queue anymore and

therefore need to change their status from "queued". While waiting for

opcode locks the job is in the "waiting" status (the constant is named

``JOB_STATUS_WAITLOCK``, but the actual value is ``waiting``). There the

following possibilities:

#. Introduce a new status, e.g. "waitdeps".

   Pro:

   - Clients know for sure a job is waiting for dependencies, not locks

   Con:

   - Code and tests would have to be updated/extended for the new status

   - List of possible state transitions certainly wouldn't get simpler

   - Breaks backwards compatibility, older clients might get confused

#. Use existing "waiting" status.

   Pro:

   - No client changes necessary, less code churn (note that there are

     clients which don't live in Ganeti core)

   - Clients don't need to know the difference between waiting for a job

     and waiting for a lock; it doesn't make a difference

   - Fewer state transitions (see commit ``5fd6b69479c0``, which removed

     many state transitions and disk writes)

   Con:

   - Not immediately visible what a job is waiting for, but it's the

     same issue with locks; this is the reason why the lock monitor

     (``gnt-debug locks``) was introduced; job dependencies can be shown

     as "locks" in the monitor

Based on these arguments, the proposal is to do the following:

- Rename ``JOB_STATUS_WAITLOCK`` constant to ``JOB_STATUS_WAITING`` to

  reflect its actual meanting: the job is waiting for something

- While waiting for dependencies and locks, jobs are in the "waiting"

  status

- Export dependency information in lock monitor; example output::

    Name      Mode Owner Pending

    job/27491 -    -     success:job/34709,job/21459

    job/21459 -    -     success,error:job/14513

Cost of deserialization

~~~~~~~~~~~~~~~~~~~~~~~

To determine the status of a dependency job the job queue must have

access to its data structure. Other queue operations already do this,

e.g. archiving, watching a job's progress and querying jobs.

Initially (Ganeti 2.0/2.1) the job queue shared the job objects

in memory and protected them using locks. Ganeti 2.2 (see :doc:`design

document <design-2.2>`) changed the queue to read and deserialize jobs

from disk. This significantly reduced locking and code complexity.

Nowadays inotify is used to wait for changes on job files when watching

a job's progress.

Reading from disk and deserializing certainly has some cost associated

with it, but it's a significantly simpler architecture than

synchronizing in memory with locks. At the stage where dependencies are

evaluated the queue lock is held in shared mode, so different workers

can read at the same time (deliberately ignoring CPython's interpreter

lock).

It is expected that the majority of executed jobs won't use

dependencies and therefore won't be affected.

Other discussed solutions

=========================

Job-level attribute

-------------------

At a first look it might seem to be better to put dependencies on

previous jobs at a job level. However, it turns out that having the

option of defining only a single opcode in a job as having such a

dependency can be useful as well. The code complexity in the job queue

is equivalent if not simpler.

Since opcodes are guaranteed to run in order, clients can just define

the dependency on the first opcode.

Another reason for the choice of an opcode-level attribute is that the

current LUXI interface for submitting jobs is a bit restricted and would

need to be changed to allow the addition of job-level attributes,

potentially requiring changes in all LUXI clients and/or breaking

backwards compatibility.

Client-side logic

-----------------

There's at least one implementation of a batched job executor twisted

into the ``burnin`` tool's code. While certainly possible, a client-side

solution should be avoided due to the different clients already in use.

For one, the :doc:`remote API <rapi>` client shouldn't import

non-standard modules. htools are written in Haskell and can't use Python

modules. A batched job executor contains quite some logic. Even if

cleanly abstracted in a (Python) library, sharing code between different

clients is difficult if not impossible.

.. vim: set textwidth=72 :

.. Local Variables:

.. mode: rst

.. fill-column: 72

.. End: