This page is a work in progress and will have more detail in the coming months. If you have specific questions, feel free to file a GitHub issue.
Materialize stores the majority of its state in memory, and works best when the streamed data can be reduced in some way. For example, if you know that only a subset of your rows and columns are relevant for your queries, it helps to avoid materializing sources or views until you’ve expressed this to the system. Materialize can then avoid stashing the full set of rows and columns, which can in some cases dramatically reduce Materialize’s memory footprint.
To prevent memory from growing without bound, Materialize periodically “compacts” data in arrangements. For example, if you have a source that tracks product inventory, you might receive periodic inventory updates throughout the day:
(T-shirts, 9:07am, +500) (T-shirts, 11:32am, -1) (T-shirts, 3:14pm, -2)
Logical compaction will fold historical updates that fall outside the compaction window into the state at the start of the window.
(T-shirts, 3:14pm, +497)
Materialize will only perform this compaction on data that falls outside the
logical compaction window. The default compaction window is 60 seconds behind
the current time, but the window can be adjusted via the
Adjusting the compaction window involves making a tradeoff between historical detail and resource usage. A larger compaction window retains more historical detail, but requires more memory. A smaller compaction window uses less memory but also retains less historical detail. Larger compaction windows also increase CPU usage, as more detailed histories require more compute time to maintain.
SELECTquery if that query depends upon multiple sources.
Note that compaction is triggered in response to updates arriving. As a result, if updates stop arriving for a source, Materialize may never compact the source fully. This can result in higher-than-expected memory usage.
This phenomenon is particularly evident when ingesting a source with a large amount of historical data (e.g, a several gigabyte Kafka topic that is no longer changing). With the default compaction window of 60 seconds, it is likely that the source will be fully ingested within the compaction window. By the time the data is eligible for compaction, the source is fully ingested, no new updates are arriving, and therefore no compaction is ever triggered.
If the increased memory usage is problematic, consider one of the following solutions:
Decrease the logical compaction window so that compaction is triggered while the source is still ingesting.
Compact the source upstream of Materialize.
If you are using the upsert envelope with a Kafka source, consider setting compaction policy on the Kafka topic to have Kafka perform the compaction.
If you are using a file source, consider rewriting the file to remove irrelevant historical data.
Periodically send dummy updates to trigger compaction.
To minimize the chances that Materialize runs out of memory in a production environment, we recommend you make additional memory available to Materialize via a SSD-backed swap file or swap partition.
This is particularly important in Linux and in Docker, where swap may not be automatically set up for you.
By default, a container has no resource
and can use as much memory and swap as the host allows, unless you have
overridden this with the
--memory or the
Most cloud Linux distributions do not enable swap by default. However, you can enable it quite easily with the following steps.
Create a swapfile using the
The general syntax is:
fallocate -l <swap size> filename. For a 1GB swap file, for example:
sudo fallocate -l 1G /swapfile
Make the swap file only accessible to
chmod 600 /swapfile
Mark the file as swap space:
Enable the swap file:
Verify the swap is available:
NAME TYPE SIZE USED PRIO /swapfile file 1G 0M -2
Optional. To make the swap file permanent, add an entry for it to
cat '/swapfile none swap sw 0 0' >> /etc/fstab
To avoid re-reading data from Kafka on restart, Materialize lets you create persistent sources, which persist input messages from Kafka topics to files on the Materialize instance’s local hard drive. The current version of persistence is not intended to speed up Materialize’s restart time, as there are other factors beyond Kafka broker read performance that contribute to high restart times.
We recommend enabling persistence if you are using Kafka sources, need to relieve load on upstream Kafka brokers, and are comfortable using experimental features.
Materialize stores one copy of all input data for each persistent Kafka source. Materialize stores these files in:
Within this directory, Materialize writes to files named
Here, each file stores data for ranges of offsets per
partition-id. Each file
stores all the data from
start-offset (inclusive) to
Materialize buffers input records in memory and flushes them every 10 minutes or
immediately if the number of buffered records per source exceeds the
--persistence-max-pending-records parameter. Setting this flag to a
higher value helps Materialize achieve higher ingest and disk write throughput,
however this also increases the average latency before records are persisted.
On restart, Materialize reads back all of the records that had been previously persisted in offset order, and then continues reading from the upstream source for data after the last persisted record in each partition.