AWX includes a powerful tool for tracking slow queries across all of its Python processes. As the AWX user, run:
$ awx-manage profile_sql --threshold 2 --minutes 5
...where threshold is the max query time in seconds, and minutes it the number of minutes to record.
For the next five minutes (in this example), any AWX Python process that generates a SQL query
that runs for >2s will be recorded in a .sqlite
database in /var/log/tower/profile
.
This is a useful tool for logging all queries at a per-process level, or filtering and searching for queries within a certain code branch. For example, if you observed that certain HTTP requests were particularly slow, you could enable profiling, perform the slow request, and then search the log:
$ sqlite3 -column -header /var/log/tower/profile/uwsgi.sqlite
sqlite> .schema queries
CREATE TABLE queries (
id INTEGER PRIMARY KEY,
version TEXT, # the AWX version
pid INTEGER, # the pid of the process
stamp DATETIME DEFAULT CURRENT_TIMESTAMP,
argv REAL, # argv of the process
time REAL, # time to run the query (in seconds)
sql TEXT, # the actual query
explain TEXT, # EXPLAIN VERBOSE ... of the query
bt TEXT # python stack trace that led to the query running
);
sqlite> SELECT time, sql FROM queries ORDER BY time DESC LIMIT 1;
time sql
---------- ---------------------------------------------------------------------------------------------
0.046 UPDATE "django_session" SET "session_data" = 'XYZ', "expire_date" = '2019-02-15T21:56:45.693290+00:00'::timestamptz WHERE "django_session"."session_key" = 'we9dumywgju4fulaxz3oki58zpxgmd6t'
Ensure that slow query logging is turned on in the Postgres config and that the log line prefix contains the application name parameter. Below is an example.
log_min_duration_statement = 500 # in ms
log_line_prefix = '< %m %a >' # timestamp, application name
We've made it easier to correlate postgres connections <--> the processes that is doing the query. For example, the following line is a log entry from a slow running awx-manage
command.
< 2021-04-21 12:42:10.307 UTC awx_1-1540765-/bin/awx-manage gather_analytics --dry-run -v 3 >LOG: duration: 1211.270 ms statement: SELECT MIN("main_jobevent"."id") AS "pk__min", MAX("main_jobevent"."id") AS "pk__max" FROM "main_jobevent" WHERE ("main_jobevent"."modified" >= '2021-03-24T12:42:08.846790+00:00'::timestamptz AND "main_jobevent"."modified" <= '2021-04-21T12:42:08.846790+00:00'::timestamptz)
The above entry was found in the log file /var/lib/pgsql/data/pg_log/postgresql-Wed.log
Note the application_name
portion. This allows us to trace the query to the node awx_1
with processes id 1540765
. The full task command line string gives us context for each long-running query that we need to find the needle in the hay stack without having to go to each individual AWX node and query Linux by the pid to understand what work is being done by each pid.
awx_1-1540765-/bin/awx-manage gather_analytics --dry-run -v 3
<tower_instance_hostname>-<pid>-<pid_launch_string>
This feature is made possible by Postgres. We do this by using the application_name
field. You can see this in pg_stat_activity
. Below is an example where we are querying pg_stat_activity
for sessions that have been alive for more than 5 minutes.
SELECT
now() - pg_stat_activity.query_start AS duration,
query,
state,
application_name
FROM pg_stat_activity
WHERE (now() - pg_stat_activity.query_start) > interval '5 minutes' and state='active';
duration | query | state | application_name
-----------------+---------------------------------------------------------------------------------------------------------------+--------+-----------------------------------------------------------------
00:13:13.125703 | COPY (SELECT main_jobevent.id, +| active | awx_1-1540765-/bin/awx-manage gather_analytics --dry-run -v 3
| main_jobevent.created, +| |
| main_jobevent.modified, +| |
| main_jobevent.uuid, +| |
| main_jobevent.parent_uuid, +| |
| main_jobevent.event, +| |
| main_jobevent.event_data::json->'task_action' AS task_action, +| |
| (CASE WHEN event = 'playbook_on_stats' THEN event_data END) as playbook_on_stats, +| |
| main_jobevent.failed, +| |
| main_jobevent.changed, +| |
| main_jobevent.playbook, +| |
| main_jobevent.play, +| |
| main_jobevent.task, +| |
| main_jobevent.role, +| |
| main_jobevent.job_id, +| |
| main_jobevent.host_id, +| |
| main_jobevent.host_name, +| |
| CAST(main_jobevent.event_data::json->>'start' AS TIMESTAMP WITH TIME ZONE) AS start,+| |
| | |
00:13:13.125703 | COPY (SELECT main_jobevent.id, +| active | awx_1-1540765-/bin/awx-manage gather_analytics --dry-run -v 3
| main_jobevent.created, +| |
| main_jobevent.modified, +| |
| main_jobevent.uuid, +| |
| main_jobevent.parent_uuid, +| |
| main_jobevent.event, +| |
| main_jobevent.event_data::json->'task_action' AS task_action, +| |
| (CASE WHEN event = 'playbook_on_stats' THEN event_data END) as playbook_on_stats, +| |
| main_jobevent.failed, +| |
| main_jobevent.changed, +| |
| main_jobevent.playbook, +| |
| main_jobevent.play, +| |
| main_jobevent.task, +| |
| main_jobevent.role, +| |
| main_jobevent.job_id, +| |
| main_jobevent.host_id, +| |
| main_jobevent.host_name, +| |
| CAST(main_jobevent.event_data::json->>'start' AS TIMESTAMP WITH TIME ZONE) AS start,+| |
| | |
Python processes in AWX's development environment are kept running in the
background via supervisord. As such, interacting with them via Python's
standard pdb.set_trace()
isn't possible.
Bundled in our container environment is a remote debugging tool, sdb
. You
can use it to set remote breakpoints in AWX code and debug interactively over
a telnet session:
# awx/main/tasks.py
class SomeTask(awx.main.tasks.jobs.BaseTask):
def run(self, pk, **kwargs):
# This will set a breakpoint and open an interactive Python
# debugger exposed on a random port between 6899-6999. The chosen
# port will be reported as a warning in the AWX logs, e.g.,
#
# [2017-01-30 22:26:04,366: WARNING/Worker-11] Remote Debugger:6900: Please telnet into 0.0.0.0 6900.
#
# You can access it from your host machine using telnet:
#
# $ telnet localhost <port>
import sdb
sdb.set_trace()
Keep in mind that when you interactively debug in this way, any process
that encounters a breakpoint will wait until an active client is established
(it won't handle additional tasks) and concludes the debugging session with
a continue
command.
To simplify remote debugging session management, AWX's development environment comes with tooling that can automatically discover open remote debugging sessions and automatically connect to them. From your host machine (i.e., outside of the development container), you can run:
sdb-listen
This will open a Python process that listens for new debugger sessions and automatically connects to them for you.
The awx-manage graph_jobs
can be used to visualize how Jobs progress from
pending to waiting to running.
awx-manage graph_jobs --help
usage: awx-manage graph_jobs [-h] [--refresh REFRESH] [--width WIDTH]
[--height HEIGHT] [--version] [-v {0,1,2,3}]
[--settings SETTINGS] [--pythonpath PYTHONPATH]
[--traceback] [--no-color] [--force-color]
Plot pending, waiting, running jobs over time on the terminal
optional arguments:
-h, --help show this help message and exit
--refresh REFRESH Time between refreshes of the graph and data in
seconds (defaults to 1.0)
--width WIDTH Width of the graph (defaults to 100)
--height HEIGHT Height of the graph (defaults to 30)
Below is an example run with 200 Jobs flowing through the system.
Decorate a function to generate profiling data that will tell you the percentage of time spent in branches of a code path. This comes at an absolute performance cost. However, the relative numbers are still very helpful.
Requirements for dot_enabled=True
Note: The profiling code will run as if dot_enabled=False
when gprof2dot
package is not found
/var/lib/awx/venv/awx/bin/pip3 install gprof2dot
Below is the signature of the @profile
decorator.
@profile(name, dest='/var/log/tower/profile', dot_enabled=True)
from awx.main.utils.profiling import profile
@profile(name="task_manager_profile")
def task_manager():
...
Now, invoke the function being profiled. Each run of the profiled function
will result in a file output to dest
containing the profile data summary as
well as a dot graph if enabled. The profile data summary can be viewed in a
text editor. The dot graph can be viewed using xdot
.
bash-4.4$ ls -aln /var/log/tower/profile/
total 24
drwxr-xr-x 2 awx root 4096 Oct 15 13:23 .
drwxrwxr-x 1 root root 4096 Oct 15 13:23 ..
-rw-r--r-- 1 awx root 635 Oct 15 13:23 2.001s-task_manager_profile-2303-272858af-3bda-45ec-af9e-7067aa86e4f3.dot
-rw-r--r-- 1 awx root 587 Oct 15 13:23 2.001s-task_manager_profile-2303-272858af-3bda-45ec-af9e-7067aa86e4f3.pstats
-rw-r--r-- 1 awx root 632 Oct 15 13:23 2.002s-task_manager_profile-2303-4cdf4660-3ef4-4238-8164-33611822d9e3.dot
-rw-r--r-- 1 awx root 587 Oct 15 13:23 2.002s-task_manager_profile-2303-4cdf4660-3ef4-4238-8164-33611822d9e3.pstats
xdot /var/log/tower/profile/2.001s-task_manager_profile-2303-272858af-3bda-45ec-af9e-7067aa86e4f3.dot
Similar to the profiling decorator, there is a timing decorator. This is useful when you do not want to incur the overhead of profiling and want to know the accurate absolute timing of a code path.
Below is the signature of the @timing
decorator.
@timing(name, dest='/var/log/tower/timing')
from awx.main.utils.profiling import timing
@timing(name="my_task_manager_timing")
def task_manager():
...
Now, invoke the function being timed. Each run of the timed function will result
in a file output to dest
. The timing data will be in the file name.
bash-4.4# ls -aln
total 16
drwxr-xr-x 2 0 0 4096 Oct 20 12:43 .
drwxrwxr-x 1 0 0 4096 Oct 20 12:43 ..
-rw-r--r-- 1 0 0 61 Oct 20 12:43 2.002178-seconds-my_task_manager-ab720a2f-4624-47d0-b897-8549fe7e8c99.time
-rw-r--r-- 1 0 0 60 Oct 20 12:43 2.00228-seconds-my_task_manager-e8a901be-9cdb-4ffc-a34a-a6bcb4266e7c.time
The class behind the decorator can also be used for profiling.
from awx.main.utils.profiling import AWXProfiler
prof = AWXProfiler("hello_world")
prof.start()
'''
code to profile here
'''
prof.stop()
# Note that start() and stop() can be reused. An new profile file will be output.
prof.start()
'''
more code to profile
'''
prof.stop()