Classic ingest patterns

You need to move away from ingest patterns commonly used for Hive 1 and 2 that are consumer centric to avoid performance problems on the consumer side.

The following diagram shows the classic partition that addresses the ingest pipeline instead of the consumer pipeline. Multiple appends to the table or partition that create small files are minimally addressed with INSERT OVERWRITE. This mitigation is not atomic.

Classic Pattern 1

This classic pattern shows repeatedly inserts: creating copy 1, copy 2, copy 3, copy 4. When you get to the sixth insert, Hive tries to write the original file, but fails because it already exists. The write attempt will fail 6 times before the insert succeeds.

This pattern can be summarized as follows:

  • Brute force.
    • Small files
    • Poor performance
  • Counter by:
    • Overwrite compaction job
    Problems:
    • Not atomic
    • Data if ingest events are not stopped

Classic pattern using partitions

When using Hive 1 or 2, creating a partition per ingest cycle helps you manage batches. This pattern results in extremely poor performance on the consumer side for the following reasons:
  • Generates many small files
  • Pressures the metastore and file system
  • Requires more compute to handle queries

Classic Pattern 2

Classic pattern 2 is similar to classic pattern 1, but adds a physical abstraction: partitioning.

Classic pattern 2 adds the pressure of partition management at the Hive metastore (HMS) level. Creating partitions per cycle, for example on day, month, year, and hour was a convenient way to manage batches for Hive 1 and 2. In Hive 3, partition ACID tables at a higher level. Instead of partitioning by day, for example, partition by month. Consider the following things:
  • How frequently you ingest the data.
  • How often you update the data.
  • The size of the data.

A daily partition thats yields 4MB of data makes sense on the ingest side, but causes problems on the consumer side. In this case, it makes sense to change the partitioning from daily to monthly. To improve the yield, partition 4MB x 30 days of data a day to yields 120MG of data. The number of small files is reduced. Hive compacts the files. Queries hit a higher density of records inside the same size ORC file.

Small partitions lead to HMS performance pressures and other problems, especially during heavy queries to the same table. Heavy reads greatly increase the load on Hive metastore (HMS) to build the partition list. If partition pruning does not occur, performance degrades. Build partitions of ACID tables on a level appropriate for your data volume. You need a huge amount of data to justify partitioning by month, day, and hour, which represents 8700 partitions per year.

Partitioning by month reduces the number of compaction operations and optimizes append operations.