In the past, I have written/seen systems where the pattern us that a task runner/worker is in charge of handling scheduling of the next tasks that need to run on completion of a task and the Scheduler only handles "issues" and initial kickoffs...
Opens another can of worms, but I think I've seen discussion on here of that idea and I thought I'd put my vote in for that pattern. On Mon, Jun 13, 2016 at 10:35 AM harish singh <[email protected]> wrote: > So I changed the scheduler heartbeat to 60 sec > > [scheduler] > job_heartbeat_sec = 5 > scheduler_heartbeat_sec = 60 > > As expected, we did see a spiky cpu utilization. > > > But I see a different problem (not so much of a problem, but just > putting it here so that it may help someone who may need to do > something similar) > > Suppose I have a DAG with Tasks as below:A-->B-->C > > If A starts and completes its execution in 10 secs. > > The next task B cannot start until the next scheduler heartbeat. That > means, I may have to wait roughly around 60 seconds in the worst case. > > So one option is to have a heartbeat of around 30 seconds > (or duration of whatever you believe is your least time consuming > task). This would just be a rough optimization to make sure we make > progress soon enough after the end of a task. > > > Thanks,Harish > > > On Mon, Jun 13, 2016 at 9:50 AM, harish singh <[email protected]> > wrote: > > > yup. it is the scheduler process using cpu: > > > > > > The below is the usage with default out-of-box heartbeat settings. > > I will paste some numbers in something, as I play with airflow.cfg. > > This is the output of `top` > > PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ > COMMAND > > * 65 airflow 20 0 445488 112648 13308 R 61.3 0.7 834:21.23 > > airflow* > > 1 airflow 20 0 17968 1912 1668 S 0.0 0.0 0:00.04 > > entrypoint.sh > > 64 airflow 20 0 407972 74876 13328 S 0.0 0.5 0:01.12 > airflow > > 154 airflow 20 0 404404 62528 4344 S 0.0 0.4 0:00.11 > airflow > > 155 airflow 20 0 404404 62532 4344 S 0.0 0.4 0:00.11 > airflow > > 156 airflow 20 0 404404 62532 4344 S 0.0 0.4 0:00.12 > airflow > > 157 airflow 20 0 404404 62532 4344 S 0.0 0.4 0:00.11 > airflow > > 158 airflow 20 0 404404 62532 4344 S 0.0 0.4 0:00.11 > airflow > > 159 airflow 20 0 404404 62532 4344 S 0.0 0.4 0:00.11 > airflow > > 160 airflow 20 0 404404 62488 4300 S 0.0 0.4 0:00.12 > airflow > > 161 airflow 20 0 404404 62468 4280 S 0.0 0.4 0:00.11 > airflow > > 163 airflow 20 0 50356 16680 5956 S 0.0 0.1 0:08.86 > > gunicorn: maste > > 168 airflow 20 0 457204 119940 13080 S 0.0 0.8 0:18.86 > > gunicorn: worke > > 170 airflow 20 0 463168 126028 13080 S 0.0 0.8 0:13.81 > > gunicorn: worke > > 171 airflow 20 0 464936 127672 13080 S 0.0 0.8 0:08.53 > > gunicorn: worke > > 174 airflow 20 0 467460 130240 13080 S 0.0 0.8 0:08.52 > > gunicorn: worke > > 7627 airflow 20 0 18208 3192 2672 S 0.0 0.0 0:00.04 bash > > 8091 airflow 20 0 30312 7556 4784 S 0.0 0.0 0:00.00 > python > > 28808 airflow 20 0 18208 3352 2836 S 0.0 0.0 0:00.08 bash > > 28822 airflow 20 0 19844 2340 2020 R 0.0 0.0 0:00.01 top > > > > > > *PID 65 is the scheduler.* > > > > > > airflow@80be4e775e55:~$ ps -ef > > UID PID PPID C STIME TTY TIME CMD > > airflow 1 0 0 Jun12 ? 00:00:00 /bin/bash ./entrypoint.sh > > airflow 64 1 0 Jun12 ? 00:00:01 /usr/bin/python > > /usr/local/bin/airflow webserver -p 8080 > > airflow 65 1 62 Jun12 ? 13:54:22 /usr/bin/python > > /usr/local/bin/airflow scheduler > > airflow 154 65 0 Jun12 ? 00:00:00 /usr/bin/python > > /usr/local/bin/airflow scheduler > > airflow 155 65 0 Jun12 ? 00:00:00 /usr/bin/python > > /usr/local/bin/airflow scheduler > > airflow 156 65 0 Jun12 ? 00:00:00 /usr/bin/python > > /usr/local/bin/airflow scheduler > > airflow 157 65 0 Jun12 ? 00:00:00 /usr/bin/python > > /usr/local/bin/airflow scheduler > > airflow 158 65 0 Jun12 ? 00:00:00 /usr/bin/python > > /usr/local/bin/airflow scheduler > > airflow 159 65 0 Jun12 ? 00:00:00 /usr/bin/python > > /usr/local/bin/airflow scheduler > > airflow 160 65 0 Jun12 ? 00:00:00 /usr/bin/python > > /usr/local/bin/airflow scheduler > > airflow 161 65 0 Jun12 ? 00:00:00 /usr/bin/python > > /usr/local/bin/airflow scheduler > > airflow 163 64 0 Jun12 ? 00:00:08 gunicorn: master > > [airflow.www.app:cached_app()] > > airflow 168 163 0 Jun12 ? 00:00:18 gunicorn: worker > > [airflow.www.app:cached_app()] > > airflow 170 163 0 Jun12 ? 00:00:13 gunicorn: worker > > [airflow.www.app:cached_app()] > > airflow 171 163 0 Jun12 ? 00:00:08 gunicorn: worker > > [airflow.www.app:cached_app()] > > airflow 174 163 0 Jun12 ? 00:00:08 gunicorn: worker > > [airflow.www.app:cached_app()] > > airflow 7627 0 0 Jun12 ? 00:00:00 bash > > airflow 8091 7627 0 Jun12 ? 00:00:00 python > > airflow 28808 0 0 16:44 ? 00:00:00 bash > > airflow 28823 28808 0 16:44 ? 00:00:00 ps -ef > > > > On Mon, Jun 13, 2016 at 8:42 AM, Maxime Beauchemin < > > [email protected]> wrote: > > > >> Can you confirm that it's the scheduler process using that CPU? > >> > >> The SCHEDULER_HEARTBEAT_SEC configuration defines a minimum duration for > >> scheduling cycles, where the scheduler evaluates all active DagRun and > >> attempts to kick off task instances whose dependencies are met. Once the > >> cycle is done, the scheduler should sleep until the next heartbeat, so > CPU > >> should look spiky. > >> > >> Max > >> > >> On Mon, Jun 13, 2016 at 8:26 AM, harish singh <[email protected] > > > >> wrote: > >> > >> > Yup, I tried changing the scheduler heartbeat to 60 seconds.. > >> > Apart from not getting any update for 60 seconds, What are the side > >> effects > >> > of changing the two heartbeats? Shouldn't impact performance? > >> > > >> > Also, I understand this cpu usage if there are 100s of dags. But with > >> just > >> > one active dag, doesnt 70% seem high? Esp in my case where there are > >> only > >> > 10 tasks in the dag making only curls (BashOperators). > >> > > >> > Also, a side now, in a different environment where we have 10 dags > >> active, > >> > the cpu usage stays in the same 70-80% range. > >> > > >> > On Mon, Jun 13, 2016, 8:14 AM Maxime Beauchemin < > >> > [email protected]> > >> > wrote: > >> > > >> > > The scheduler constantly attempts to schedule tasks, interacting > with > >> the > >> > > database and reloading DAG definition. In most larg-ish > environments, > >> > > burning up to a CPU to run the scheduler doesn't seem outrageous to > >> me. > >> > > > >> > > If you want to reduce the CPU load related to the scheduler check > out > >> > > SCHEDULER_HEARTBEAT_SEC and MAX_THREADS in the scheduler section of > >> > > `airflow.cfg` > >> > > > >> > > Max > >> > > > >> > > On Sun, Jun 12, 2016 at 1:24 PM, harish singh < > >> [email protected]> > >> > > wrote: > >> > > > >> > > > Hi guys, > >> > > > > >> > > > We are running airflow (for about 3 months now) inside a docker > >> > container > >> > > > on aws. > >> > > > > >> > > > I just did a docker stats to check whats going on. The cpu > >> consumption > >> > is > >> > > > huge. > >> > > > We have around 15 DAGS. Only one DAG is turned ON. the remaining > are > >> > OFF. > >> > > > The DAG runs with a HOURLY schedule. > >> > > > > >> > > > Right now, airflow is consuming almost 1 complete core. > >> > > > It seems there is some unnecessary spinning? > >> > > > This doesnt look like the right behavior. > >> > > > Is there a bug already filed for this? Or am not sure if there is > >> > > something > >> > > > incorrect in the way I am using the airflow configuration. > >> > > > > >> > > > CONTAINER CPU % MEM USAGE / LIMIT MEM > % > >> > > > NET I/O BLOCK I/O > >> > > > CCC 68.17% 619.7 MB / 2.147 > GB > >> > > > 28.85% 1.408 GB / 939.4 MB 7.856 MB / 0 B > >> > > > XXX 64.36% 619.4 MB / 2.147 > GB > >> > > > 28.84% 1.211 GB / 807.6 MB 7.856 MB / 0 B > >> > > > > >> > > > > >> > > > Ariflow version 1.7.0 > >> > > > > >> > > > Airflow config: > >> > > > > >> > > > sql_alchemy_pool_size = 5 > >> > > > sql_alchemy_pool_recycle = 3600 > >> > > > parallelism = 8 > >> > > > dag_concurrency = 8 > >> > > > max_active_runs_per_dag = 8 > >> > > > > >> > > > > >> > > > > >> > > > Thanks, > >> > > > > >> > > > Harish > >> > > > > >> > > > >> > > >> > > > > >
