What they are¶
A pillow is a subscriber to a change feed. When a change is published the pillow receives the document, performs some calculation or transform, and publishes it to another database.
Creating a pillow¶
All pillows inherit from ConstructedPillow class. A pillow consists of a few parts:
- Change Feed
- Change Event Handler
Change feeds are documented here.
The 10,000 foot view is a change feed publishes changes which you can subscribe to.
The checkpoint is a json field that tells processor where to start the change feed.
A processor is what handles the transformation or calculation and publishes it to a database. Most pillows only have one processor, but sometimes it will make sense to combine processors into one pillow when you are only iterating over a small number of documents (such as custom reports).
When creating a processor you should be aware of how much time it will take to process the record. A useful baseline is:
86400 seconds per day / # of expected changes per day = how long your processor should take
Note that it should be faster than this as most changes will come in at once instead of evenly distributed throughout the day.
Change Event Handler¶
This fires after each change has been processed. The main use case is to save the checkpoint to the database.
Pillow errors are handled by saving to model PillowError. A celery queue reads from this model and retries any errors on the pillow.
There are several datadog metrics with the prefix commcare.change_feed that can be helpful for monitoring pillows.
For UCR pillows the pillow log will contain any data sources and docs that have exceeded a threshold and can be used to find expensive data sources.
A pillow is falling behind¶
A pillow can fall behind for two reasons:
- The processor is too slow for the number of changes that are coming in.
- There has been an issue with the change feed that has caused the checkpoint to be “rewound”
Optimizing a processor¶
To solve #1 you should use any monitors that have been set up to attempt to pinpoint the issue.
If this is a UCR pillow use the profile_data_source management command to profile the expensive data sources.
To scale UCR Pillows horizontally do the following:
- Look for what pillows are behind. This can be found in the change feed dashboard or the hq admin system info page.
- Ensure you have enough resources on the pillow server to scale the pillows This can be found through datadog.
- Decide what topics need to have added partitions in kafka. There is no way to scale a couch pillow horizontally. You can also not remove partitions so you should attempt scaling in small increments. Also attempt to make sure pillows are able to split partitions easily. It’s easiest to use powers of 2
- Run ./manage.py add_kafka_partition <topic> <number partitions to have>
- In the commcare-hq-deploy repo change num_processes to the pillows you want to scale. Merge this PR and update the submodule in hq
- On the next deploy multiple processes will be used when starting pillows
The UCR pillows also have options to split the pillow into multiple. They include ucr_divsion, include_ucrs and exclude_ucrs. Look to the pillow code for more information on these.
Occasionally checkpoints will be “rewound” to a previous state causing pillows to process changes that have already been processed. This usually happens when a couch node fails over to another. If this occurs, stop the pillow, wait for confirmation that the couch nodes are up, and fix the checkpoint using: ./manage.py fix_checkpoint_after_rewind <pillow_name>
Problem with checkpoint for pillow name: First available topic offset for topic is num1 but needed num2¶
This happens when the earliest checkpoint that kafka knows about for a topic is after the checkpoint the pillow wants to start at. This often happens if a pillow has been stopped for a month and has not been removed from the settings.
To fix this you should verify that the pillow is no longer needed in the environment. If it isn’t, you can delete the checkpoint and re-deploy. This should eventually be followed up by removing the pillow from the settings.
If the pillow is needed and should be running you’re in a bit of a pickle. This means that the pillow is not able to get the required document ids from kafka. It also won’t be clear what documents the pillows has and has not processed. To fix this the safest thing will be to force the pillow to go through all relevant docs. Once this process is started you can move the checkpoint for that pillow to the most recent offset for its topic.