Parallel execution¶
New in version 1.3.
By default, Fabric executes all specified tasks serially (see Execution strategy for details.) This document describes Fabric’s options for running tasks on multiple hosts in parallel, via per-task decorators and/or global command-line switches.
What it does¶
Because Fabric 1.x is not fully threadsafe (and because in general use, task functions do not typically interact with one another) this functionality is implemented via the Python multiprocessing module. It creates one new process for each host and task combination, optionally using a (configurable) sliding window to prevent too many processes from running at the same time.
For example, imagine a scenario where you want to update Web application code on a number of Web servers, and then reload the servers once the code has been distributed everywhere (to allow for easier rollback if code updates fail.) One could implement this with the following fabfile:
from fabric.api import *
def update():
with cd("/srv/django/myapp"):
run("git pull")
def reload():
sudo("service apache2 reload")
and execute it on a set of 3 servers, in serial, like so:
$ fab -H web1,web2,web3 update reload
Normally, without any parallel execution options activated, Fabric would run in order:
update
onweb1
update
onweb2
update
onweb3
reload
onweb1
reload
onweb2
reload
onweb3
With parallel execution activated (via -P
– see below for details),
this turns into:
update
onweb1
,web2
, andweb3
reload
onweb1
,web2
, andweb3
Hopefully the benefits of this are obvious – if update
took 5 seconds to
run and reload
took 2 seconds, serial execution takes (5+2)*3 = 21 seconds
to run, while parallel execution takes only a third of the time, (5+2) = 7
seconds on average.
How to use it¶
Decorators¶
Since the minimum “unit” that parallel execution affects is a task, the
functionality may be enabled or disabled on a task-by-task basis using the
parallel
and serial
decorators. For
example, this fabfile:
from fabric.api import *
@parallel
def runs_in_parallel():
pass
def runs_serially():
pass
when run in this manner:
$ fab -H host1,host2,host3 runs_in_parallel runs_serially
will result in the following execution sequence:
runs_in_parallel
onhost1
,host2
, andhost3
runs_serially
onhost1
runs_serially
onhost2
runs_serially
onhost3
Command-line flags¶
One may also force all tasks to run in parallel by using the command-line flag
-P
or the env variable env.parallel. However,
any task specifically wrapped with serial
will ignore this
setting and continue to run serially.
For example, the following fabfile will result in the same execution sequence as the one above:
from fabric.api import *
def runs_in_parallel():
pass
@serial
def runs_serially():
pass
when invoked like so:
$ fab -H host1,host2,host3 -P runs_in_parallel runs_serially
As before, runs_in_parallel
will run in parallel, and runs_serially
in
sequence.
Bubble size¶
With large host lists, a user’s local machine can get overwhelmed by running too many concurrent Fabric processes. Because of this, you may opt to use a moving bubble approach that limits Fabric to a specific number of concurrently active processes.
By default, no bubble is used and all hosts are run in one concurrent pool. You
can override this on a per-task level by specifying the pool_size
keyword
argument to parallel
, or globally via -z
.
For example, to run on 5 hosts at a time:
from fabric.api import *
@parallel(pool_size=5)
def heavy_task():
# lots of heavy local lifting or lots of IO here
Or skip the pool_size
kwarg and instead:
$ fab -P -z 5 heavy_task
Linewise vs bytewise output¶
Fabric’s default mode of printing to the terminal is byte-by-byte, in order to support Interaction with remote programs. This often gives poor results when running in parallel mode, as the multiple processes may write to your terminal’s standard out stream simultaneously.
To help offset this problem, Fabric’s option for linewise output is automatically enabled whenever parallelism is active. This will cause you to lose most of the benefits outlined in the above link Fabric’s remote interactivity features, but as those do not map well to parallel invocations, it’s typically a fair trade.
There’s no way to avoid the multiple processes mixing up on a line-by-line basis, but you will at least be able to tell them apart by the host-string line prefix.
Note
Future versions will add improved logging support to make troubleshooting parallel runs easier.