Peter, The numbers you mention seem to come out of the blue. I think you’re oversimplifying it and cannot simply state 180/36 = 5 minutes. Throwing in numbers without explanation creates confusion.
I have some questions when reading your AIP. I have to make lots of assumptions and think explaining it more in depth would clarify a lot. * Should a DagScheduler and task run in the same slots? Should there be any difference between running the two? * How does this work together with e.g. Celery? With the current Celery setup you push tasks to the queue, which can be run on any worker. With your setup it seems you push a DagScheduler and/or task. Does the DagScheduler push tasks itself to the queue so that any task can run anywhere? And is it correct to assume the DagScheduler polls the DB to check if a task is finished? * “If a cycle is done the MainScheduler should schedule a new DagScheduler" -> I assume the dagscheduler would set state of a dagrun to finished? And the mainscheduler simply checks when the next interval is finished and to start a new DagScheduler. * The naming is confusing to me. How about naming the DagScheduler “DagRunManager”, because that’s what I believe it does? * I’m not convinced this is the way to go. Currently the scheduler process does a lot more than just scheduling. I.e. also parsing of the DAGs, which I believe can be optimised a lot. I think splitting up the responsibilities of the scheduler would be more beneficial, instead of adding complexity by creating more “schedulers”. Would you agree? On a final note, to this whole thread: I’m very wary of doing HA/distributed work ourselves. It adds a lot of complexity, locking is a tricky subject (although well thought out by many others, so not impossible). Before going there, I would suggest to put effort into optimising the (single) scheduler first, e.g. by splitting responsibilities between DAG parsing and actual scheduling. Cheers, Bas On 18 Mar 2019, at 07:18, Peter van t Hof <pjrvant...@gmail.com<mailto:pjrvant...@gmail.com>> wrote: Hi, My proposal is focusing mainly on scalability and indeed not so much on HA. This mainly because that is also the main issue from the original author. Have a form of HA on this MainScheduler would still be nice to have. The problem with is that have a fixed number of scheduler does not scale on the load. On my current client they try to execute 5000+ DAG’s at the same time. A single scheduler cycle to touch all DAG’s takes 2-3 hour. So to do this within 5 min 36 of those schedulers with locking should be there at all time. After 2 hours 2 schedulers would be enough, this means in this situation 34 scheduler processes are wasted and only producing overhead. This DagScheduler is a short living task, so this is not a persistent worker process. The MainScheduler should resubmit when it is required. Gr, Peter On 18 Mar 2019, at 05:32, Maxime Beauchemin <maximebeauche...@gmail.com<mailto:maximebeauche...@gmail.com>> wrote: The proposal reads "Looking at the original AIP-15 the author proposes to use locking to enable the use of multiple schedulers, this might introduce unnecessary complexity" To me introducing multiple roles (master scheduler + scheduler minions), may be actually more complex than just having "shared nothing" schedulers with locking. The former is also less scalable (whatever workload is done on that master [say serialization] can hit scale issues) and is less HA (as it relies on the orchestrator [k8s] for HA). My personal incline has always been going towards renaming the scheduler to "supervisor" (as it already does significantly more than just triggering tasks) and allowing many instances of that role, and using locks where necessary. That way there are just 2 roles in the cluster: supervisor and worker processes. Depending on the executor (say for k8s) you don't even need actual persistent worker processes. Max On Sun, Mar 17, 2019 at 1:52 AM Peter van t Hof <pjrvant...@gmail.com<mailto:pjrvant...@gmail.com>> wrote: Hi all, I think that scheduling locking is maybe not the best way in solving this issue. Still I’m in support of taking a good look at the scheduler because it has some real scaling issues. I did wrote an alternative proposal to solve the scalability of the scheduler: https://cwiki.apache.org/confluence/display/AIRFLOW/AIP-15+Scalable+Scheduler < https://cwiki.apache.org/confluence/display/AIRFLOW/AIP-15+Scalable+Scheduler Any input on this is welcome. Gr, Peter On 3 Mar 2019, at 03:26, Deng Xiaodong <xd.den...@gmail.com> wrote: Thanks Max. I have documented all the discussions around this topic & useful inputs into AIP-15 (Support Multiple-Schedulers for HA & Better Scheduling Performance) https://cwiki.apache.org/confluence/pages/viewpage.action?pageId=103092651 < https://cwiki.apache.org/confluence/pages/viewpage.action?pageId=103092651>. More inputs from folks are welcomed. Thanks. XD On 3 Mar 2019, at 6:18 AM, Maxime Beauchemin < maximebeauche...@gmail.com> wrote: Personally I'd vote against the idea of having certain scheduler handling a subset of the DAGs, that's just not HA. Also if you are in an env where you have a small number of large DAGs, the odds of having wasted work and double-firing get pretty high. With the lock in place, it's just a matter of the scheduler loop to select (in a db transaction) the dag that's not been processed for the longest time that is not locked. Flipping the lock flag to true should be part of the db transaction. We probably need a btree index on lock and last processed time. This way adding scheduler processes increases the scheduling pace, and provides an HA solution. No leader / master / slave or election process, just equal workers that work together. Max On Sat, Mar 2, 2019 at 7:04 AM Deng Xiaodong <xd.den...@gmail.com> wrote: Get your point and agree. And the suggestion you gave lastly to random sort DAGs is a great idea to address it. Thanks! XD On 2 Mar 2019, at 10:41 PM, Jarek Potiuk <jarek.pot...@polidea.com> wrote: I think that the probability calculation holds only if there is no correlation between different schedulers. I think however there might be an accidental correlation if you think about typical deployments. Some details why I think accidental correlation is possible and even likely. Assume that: - we have similar and similarly busy machines running schedulers (likely) - time is synchronised between the machines (likely) - the machines have the same DAG folders mounted (or copied) and the same filesystem is used (this is exactly what multiple schedulers deployment is all about) - the schedulers start scanning at exactly the same time (crossing 0:00 second every full five minutes for example) - this I am not sure but I imagine this might be "typical" behaviour. - they process list of DAGs in exactly the same sequence (it looks like this is the case dag_processing < https://github.com/apache/airflow/blob/master/airflow/utils/dag_processing.py#L300 and models/__init__ < https://github.com/apache/airflow/blob/master/airflow/models/__init__.py#L567 : we use os.walk which uses os.listdir for which sequence of processing depends on the filesystem implementation < https://stackoverflow.com/questions/31534583/is-os-listdir-deterministic> and then we append files to the list) Then it's rather likely that the schedulers will be competing about the very same DAGs at the very beginning. Locking will change how quickly they process each DAG of course, but If the DAGs are of similar sizes it's also likely that the speed of scanning (DAGS/s) is similar for all schedulers. The schedulers will then catch-up with each other and might pretty much continuously compete for the same DAGs almost all the time. It can be mitigated super-easily by random sorting of the DAGs folder list after it is prepared (it's file-system dependent now so we do not rely on particular order) . Then the probability numbers will hold perfectly I think :) J. On Sat, Mar 2, 2019 at 2:41 PM Deng Xiaodong <xd.den...@gmail.com> wrote: I’m thinking of which architecture would be ideal. # Option-1: The master-slave architecture would be one option. But leader-selection will be very essential to consider, otherwise we have issue in terms of HA again. # Option-2: Another option we may consider is to simply start multiple scheduler instances (just using the current implementation, after modify & validate the scheduler_lock on DagModel). - In this case, given we handle everything properly using locking, we don’t need to worry too much about double-scheduling/triggering. - Another potential concern I had earlier is that different schedulers may compete with each other and cause “waste” of scheduler resource. After further thinking, I realise this is a typical Birthday Problem. Given we have m DAGs, and n schedulers, at any moment, the probability that all schedulers are working on different DAGs is m!/((m-n)! * (m^n)), and the probability that there are schedulers competing on the same DAG will be 1-m!/((m-n)! * (m^n)). Let’s say we have 200 DAGs and we start 2 schedulers. At any moment, the probability that there is schedulers competing on the same DAG is only 0.5%. If we run 2 schedulers against 300 DAGs, this probability is only 0.33%. (This probability will be higher if m/n is low. But users should not start too many schedulers if they don’t have that many DAGs). Given the probability of schedulers competing is so low, my concern on scheduler resource waste is not really valid. Based on these calculations/assessment, I think we can go for option-2, i.e. we don’t make big change in the current implementation. Instead, we ensure the scheduler_lock is working well and test intensively on running multiple schedulers. Then we should be good to let users know that it’s safe to run multiple schedulers. Please share your thoughts on this and correct me if I’m wrong in any point above. Thanks. XD Reference: https://en.wikipedia.org/wiki/Birthday_problem < https://en.wikipedia.org/wiki/Birthday_problem> On 2 Mar 2019, at 3:39 PM, Tao Feng <fengta...@gmail.com> wrote: Does the proposal use master-slave architecture(leader scheduler vs slave scheduler)? On Fri, Mar 1, 2019 at 5:32 PM Kevin Yang <yrql...@gmail.com> wrote: Preventing double-triggering by separating DAG files different schedulers parse sounds easier and more intuitive. I actually removed one of the double-triggering prevention logic here < https://github.com/apache/airflow/pull/4234/files#diff-a7f584b9502a6dd19987db41a8834ff9L127 (expensive) and was relying on this lock < https://github.com/apache/airflow/blob/master/airflow/models/__init__.py#L1233 to prevent double-firing and safe-guard our non-idempotent tasks( btw the insert can be insert overwrite to be idempotent). Also tho in Airbnb we requeue tasks a lot, we haven't see double-firing recently. Cheers, Kevin Y On Fri, Mar 1, 2019 at 2:08 PM Maxime Beauchemin < maximebeauche...@gmail.com> wrote: Forgot to mention: the intention was to use the lock, but I never personally got to do the second phase which would consist of skipping the DAG if the lock is on, and expire the lock eventually based on a config setting. Max On Fri, Mar 1, 2019 at 1:57 PM Maxime Beauchemin < maximebeauche...@gmail.com> wrote: My original intention with the lock was preventing "double-triggering" of task (triggering refers to the scheduler putting the message in the queue). Airflow now has good "double-firing-prevention" of tasks (firing happens when the worker receives the message and starts the task), even if the scheduler was to go rogue or restart and send multiple triggers for a task instance, the worker(s) should only start one task instance. That's done by running the database assertions behind the conditions being met as read database transaction (no task can alter the rows that validate the assertion while it's getting asserted). In practice it's a little tricky and we've seen rogue double-firing in the past (I have no idea how often that happens). If we do want to prevent double-triggerring, we should make sure that 2 schedulers aren't processing the same DAG or DagRun at the same time. That would mean for the scheduler to not start the process of locked DAGs, and by providing a mechanism to expire the locks after some time. Has anyone experienced double firing lately? If that exist we should fix it, but also be careful around multiple scheduler double-triggering as it would make that problem potentially much worse. Max On Fri, Mar 1, 2019 at 8:19 AM Deng Xiaodong < xd.den...@gmail.com> wrote: It’s exactly what my team is doing & what I shared here earlier last year ( https://lists.apache.org/thread.html/0e21230e08f07ef6f8e3c59887e9005447d6932639d3ce16a103078f@%3Cdev.airflow.apache.org%3E < https://lists.apache.org/thread.html/0e21230e08f07ef6f8e3c59887e9005447d6932639d3ce16a103078f@%3Cdev.airflow.apache.org%3E ) It’s somehow a “hacky” solution (and HA is not addressed), and now I’m thinking how we can have it more proper & robust. XD On 2 Mar 2019, at 12:04 AM, Mario Urquizo < mario.urqu...@gmail.com> wrote: We have been running multiple schedulers for about 3 months. We created multiple services to run airflow schedulers. The only difference is that we have each of the schedulers pointed to a directory one level deeper than the DAG home directory that the workers and webapp use. We have seen much better scheduling performance but this does not yet help with HA. DAGS_HOME: {airflow_home}/dags (webapp & workers) {airflow_home}/dags/group-a/ (scheduler1) {airflow_home}/dags/group-b/ (scheduler2) {airflow_home}/dags/group-etc/ (scheduler3) Not sure if this helps, just sharing in case it does. Thank you, Mario On Fri, Mar 1, 2019 at 9:44 AM Bolke de Bruin < bdbr...@gmail.com> wrote: I have done quite some work on making it possible to run multiple schedulers at the same time. At the moment I don’t think there are real blockers actually to do so. We just don’t actively test it. Database locking is mostly in place (DagRuns and TaskInstances). And I think the worst that can happen is that a task is scheduled twice. The task will detect this most of the time and kill one off if concurrent if not sequential then I will run again in some occasions. Everyone is having idempotent tasks right so no harm done? ;-) Have you encountered issues? Maybe work those out? Cheers Bolke. Verstuurd vanaf mijn iPad Op 1 mrt. 2019 om 16:25 heeft Deng Xiaodong < xd.den...@gmail.com> het volgende geschreven: Hi Max, Following https://lists.apache.org/thread.html/0e21230e08f07ef6f8e3c59887e9005447d6932639d3ce16a103078f@%3Cdev.airflow.apache.org%3E < https://lists.apache.org/thread.html/0e21230e08f07ef6f8e3c59887e9005447d6932639d3ce16a103078f@%3Cdev.airflow.apache.org%3E , I’m trying to prepare an AIP for supporting multiple-scheduler in Airflow (mainly for HA and Higher scheduling performance). Along the process of code checking, I found that there is one attribute of DagModel, “scheduler_lock”. It’s not used at all in current implementation, but it was introduced long time back (2015) to allow multiple schedulers to work together ( https://github.com/apache/airflow/commit/2070bfc50b5aa038301519ef7c630f2fcb569620 < https://github.com/apache/airflow/commit/2070bfc50b5aa038301519ef7c630f2fcb569620 ). Since you were the original author of it, it would be very helpful if you can kindly share why the multiple-schedulers implementation was removed eventually, and what challenges/complexity there were. (You already shared a few valuable inputs in the earlier discussion https://lists.apache.org/thread.html/d37befd6f04dbdbfd2a2d41722352603bc2e2f97fb47bdc5ba454d0c@%3Cdev.airflow.apache.org%3E < https://lists.apache.org/thread.html/d37befd6f04dbdbfd2a2d41722352603bc2e2f97fb47bdc5ba454d0c@%3Cdev.airflow.apache.org%3E , mainly relating to hiccups around concurrency, cross DAG prioritisation & load on DB. Other than these, anything else you would like to advise?) I will also dive into the git history further to understand it better. Thanks. XD -- Jarek Potiuk Polidea <https://www.polidea.com/> | Principal Software Engineer M: +48 660 796 129 <+48660796129> E: jarek.pot...@polidea.com
