A problem that I have run into repeatedly when doing schema design is how to control partition size while still allowing for efficient multi-row queries.
We want to limit partition size to some number between 10 and 100 megabytes to avoid operational issues. The standard way to do that is to figure out the maximum number of rows that your "natural partition key" will ever need to support and then add an additional artificial partition key that segments the rows sufficiently to get keep the partition size under the maximum. In the case of time series data, this is often done by bucketing by time period, i.e. creating a new partition every minute, hour or day. For non-time series data by doing something like Hash(clustering-key) mod desired-number-of-partitions. In my case, multi-row queries to support a REST API typically return a page of results, where the page size might be anywhere from a few dozen up to thousands. For query efficiency I want the average number of rows per partition to be large enough that a query can be satisfied by reading a small number of partitions--ideally one. So I want to simultaneously limit the maximum number of rows per partition and yet maintain a large enough average number of rows per partition to make my queries efficient. But with my data the ratio between maximum and average can be very large (up to four orders of magnitude). Here is an example: Rows per Partition Partition Size Mode 1 1 KB Median 500 500 KB 90th percentile 5,000 5 MB 99th percentile 50,000 50 MB Maximum 2,500,000 2.5 GB In this case, 99% of my data could fit in a single 50 MB partition. But if I use the standard approach, I have to split my partitions into 50 pieces to accommodate the largest data. That means that to query the 700 rows for my median case, I have to read 50 partitions instead of one. If you try to deal with this by starting a new partition when an old one fills up, you have a nasty distributed consensus problem, along with read-before-write. Cassandra LWT wasn't available the last time I dealt with this, but might help with the consensus part today. But there are still some nasty corner cases. I have some thoughts on other ways to solve this, but they all have drawbacks. So I thought I'd ask here and hope that someone has a better approach. Thanks in advance, Jim
