A previous article in the PostgreSQL Concurrency series covered how to manage concurrent retweets in an efficient way: in Computing and Caching, we learnt how to maintain a cache right in your PostgreSQL database, using MATERIALIZED VIEWS. We also had a look at how to take care of Batch Updates and Concurrency.

While in the first case we are providing a solution to a technical problem where we want to solve performance issues while keeping the same semantics, in the second case we are actually implementing a part of the application’s Business Logic as a scheduled job.

Today’s article shows a modern technique to handle the scheduling of those business oriented activities that are not tied to any user activity. When thinking about it this way, you certainly don’t want to implement the backbone of your business logic in a shell script that’s directly maintained in the production environment, do you?

Scheduling Data Processing

Data processing is most often triggered by user actions, either your end-user or internal users: back-office, admin superusers, or support users maybe. In some cases though, it may happen that some data processing needs to happen on its own, following a schedule rather than being triggered by direct user activity on your product.

As we saw in Computing and Caching, the use case might be a technical implementation detail. It might also be user oriented, like daily dashboard metrics computating, account consolidating, user-defined alerting schedule (daily or weekly summaries, etc), or many other use cases really.

When the use case for scheduling data processing is business oriented, then the usual questions about SQL and Business Logic are back in play again. If we’re implementing daily activity reporting by email to our users, of course the email contents are going to be driven by business logic. The numbers you’re sending are the result of the business rules used to compute them.

So we have to trigger data processing at fixed moments in time rather than triggering it with some user activity. The classic way to implement that is to use the venerable crontab facility from Unix systems.

In classic Unix phylosophy, Cron is great at just one thing: starting a user defined command at a specific time specification, which can be… almost anything. Once in a blue moon. Every other Thursday at 6pm. Every morning. At boot time. About any point in a calendar can be specified as a cron schedule expression, and that includes specifications of recurring events.

How NOT to use cron?

The classic approach to cron jobs could be qualified as surprising, if you want to be nice. It’s actually wrong in that it solves none of the interesting problems you have when using cron.

A very classic cron usage looks like the following:

0 22 * * 1-5 /usr/local/bin/my-script.sh

The cron unix service is then going to call /usr/local/bin/my-script.sh at 22:00 on every day-of-week from Monday through Friday. All that’s left is for the bash script to implement the data processing facility we need.

An improvement to the bash script that is often found in production environments consists of using a reasonable programming language to write your cron command in. By that I mean a programming language that makes it easy to handle unforeseen situations: it usually requires a kind of a condition system or at least the ability to handle errors and exceptions.

Your bash script might being with a call to trap if you’re an advanced enough user, unfortunately that’s pretty rare and only gets you so far.

Even when using a reasonable programming language, the needs for a processing job that is scheduled in the background are a little more involved than that.

Code Management and Deployment

Another important aspect in that usage of cron is the code management. Often enough cron commands are bash scripts. Sometimes cron commands are properly implemented with the same programming language as the rest of the user facing application, using the same business rules and code.

In both cases, it’s important to note that the code should be managed in a proper code repository, versionned, and deployed with the same process and quality as applied to the rest of the code.

In many places, even when people are serious about their deployments, I’ve seen cron jobs falling in between the devs and ops team, no one being the owner of them, and thus no clear procedure would be in place for maintaining those scripts.

The Real Needs of a Scheduled Data Processing Job

So, why don’t I just run my cron jobs in bash, you’re asking? Well because here’s a list of things that I want to be able to easily know, verify and act on when dealing with a data processing job scheduled in the background:

  • When did it run? for how long?

    I know when I said it should start, that’s the cron specification. Of course I might have been wrong when writting that, so I want to be able to debug it easily. Also, we might have a system that prevents the same processing job to run more than once on the same server, usually that would be achieved with a lock file.

    A very classic gotcha is to forget about how long the job is taking when computing its frequency. Say we want to run a job every 5 minutes, so we have a cron entry like the following:

    */5 * * * * /usr/local/bin/my-script.sh

    Then if we have a lockfile (we usually should) and the processing runs for more than five minutes, we’re going to skip scheduled runs. If the aim is to guarantee that the data is at least as fresh as 5 minutes ago, that’s a problem and I want to know about it.

    Sure, I can send messages to the logs and have some automated processing for that, and most cron/bash users probably use logger for that already, with a complex log processing setup behind it (logstash or ELK are some of the popular stacks for handling logs).

  • What did the job had to do, what did it do?

    Now, as soon as the job is implementing some business logic, I want to have some metrics about what the data processing job did, in terms of my business logic, and I want those metrics displayed at a place where our internal teams (accounting, finance, marketing, etc) can see them, and expressed in a way that makes sense for them.

  • How to check if the processing was correct?

    Depending on what the processing is all about, it might be worthwile to have both a summary view of what the job did (how many items it processed, to which cumulated amount of value, etc), and sometimes a detailed view where you can dive in any of the processed items and see what the processing was all about.

    If you’re using PostgreSQL, then having a generic before/after diff is easy enough to implement thanks to the hstore extension, as detailed in my article Auditing Changes with Hstore.

  • Was there any failure? how to review and process the failures?

    Now if some items failed to be processed, our job might be smart enough to mark them as failed and continue with other tasks. In that case, someone has to review the failed items and do something about them.

    Or maybe it’s better to fail early and just retry that same job over and over again, assuming that something will change that allow the processing to succeed. An upgrade of the data processing utility, for instance.

    It might be that your application is now producing new kinds of events that you don’t know how to handle yet. Or just a plain bug that you’re lucky enough is failing the processing rather than silently producing the wrong results…

  • If a job is still running, can I stop it? restart it? cancel it? reschedule it?

    And that one is pretty important. Say you have a nightly job to run and for some reasons it failed to run yesternight. Maybe because you scheduled it at 2:30am but thanks to daylight saving this time specification didn’t happen yesterday. Yes, that’s a classic. So what now? Can you run the cron job again easily in production?

    Another use case here is that famous infinite loop that only happens with the exact right setup and data combo, and so you didn’t think of guarding against the case in your code because it won’t happen. Well, it did. It’s actually happening now. How easy is it for you to take control of your background job?

Those are some pretty basic questions here, and if the data processing jobs that you schedule are important for your business, then I’m sure you want to be able to answer all of them.

Not all cron jobs are the same

It must also be said that we have two very different cases to think about when discussing cron jobs:

  1. Technical cron jobs, like archiving, backups and restore, and other commands related to the operational architecture.

    In that case, most probably the team that owns the cron commands is an ops team, and the tooling they use day in and day out is going to be well suited for making all the previous requirements easy enough to implement for them.

    After all, ssh, ps, kill, rm -f and all are proper tooling if you’re an ops guy debugging something unexpected in production, in many cases. And when you prefer more automation around your ops process, I’m sure you know how to include cron jobs debugging in your tooling.

  2. Business Logic that needs to be scheduled at known time specification.

    That’s what this blog article is all about. Because now, you need to write application code that happens to run when specified by your crontab, or with some other scheduler.

    Now debugging the job isn’t going to be done by the ops team, because they’re not tasked with understanding the fine points of the business rules, and most of the bugs you’ll want to address here aren’t going to be related to your production’s architecture, rather to those business rules.

    That’s when the tooling used in production should be easy enough for an application developer perspective. And in most places, having to fiddle around interactively on production systems to fix a business rule implementation bug isn’t the best way to ensure quality.

We need a solution that can enable a developer team to manage the scheduled events properly, and without requiring fiddling interactively at a production’s system shell.

Proper Usage of Cron

The idea that I want to push forward in this article is the following: when you need to schedule data processing, write a web service that implements your data processing, with full manual controls over it, a dashboard and logs. Some of the things your service must implement are:

  • A dashboard to know what’s been done when, and what’s currently happening.

  • An easy way to read the logs from your background activities.

  • Some ways to see the failed processing.

  • Maybe a way to manually handle the failures, or re-inject them into the pipeline after having (maybe?) fixed exposed bugs.

  • A control to start, stop, resume, and restart current jobs.

  • An API that allows to start a new job non-interactively, and that we are going to use in our crontab entry.

One way to get started is to implement those features on top of a very basic data processing function: a function that fails with a Not Yet Implemented error. Then you can debug your background job by adding the missing rules and processing, and use the interactive controls to restart job processing from the application, and see the dashboards getting updated with success and failures, and metrics that make sense for whatever it is you are implementing.

When you have it working, it’s time to deploy the service and implement the scheduling parts. As your background data processing jobs are now implemented as web services, your organisation probably already knows how to deploy it, version it, do maintenance upgrades and bug fixes, etc.

About scheduling the jobs, then use cron to produce events in your system, calling a service URL that triggers the processing of the next batch of data:

0 22 * * 1-5 curl https://service.internal.url/schedule/daily-reporting/
*/5 * * * *  curl https://matviews.internal.url/schedule/refresh-mat-views/

You can then browse to your interactive web application and see the dashboard with the interesting metrics, a summary of what happened when, how many users got a daily reporting, maybe even have a copy of the sent email, etc.

Conclusion

Some cron jobs are technical details meant for sysadmins, and then using system logs to track the activity is fair enough. Some cron jobs are implementing a user visible part of your business logic, and as such they need to expose business metrics and allow for direct control of the running tasks, their scheduling, and the processing itself.

The best way to achieve proper business logic background jobs scheduling is to actually write a fully interactive server-side application (a web app, typically), and have it expose an API call that inject a scheduler event in the application.

When doing things that way, then cron is used for what it’s good at: producing an event at known regular timings, controled by the calendar.