Data Distribution
This documentation is for an unreleased version of Apache Paimon. We recommend you use the latest stable version.

Data Distribution #

By default, Paimon table only has one bucket, which means it only provides single parallelism read and write. Please configure the bucket strategy to your table.

A bucket is the smallest storage unit for reads and writes, each bucket directory contains an LSM tree.

Fixed Bucket #

Configure a bucket greater than 0, using Fixed Bucket mode, according to Math.abs(key_hashcode % numBuckets) to compute the bucket of record.

Rescaling buckets can only be done through offline processes, see Rescale Bucket. A too large number of buckets leads to too many small files, and a too small number of buckets leads to poor write performance.

Dynamic Bucket #

Configure 'bucket' = '-1'. The keys that arrive first will fall into the old buckets, and the new keys will fall into the new buckets, the distribution of buckets and keys depends on the order in which the data arrives. Paimon maintains an index to determine which key corresponds to which bucket.

Paimon will automatically expand the number of buckets.

  • Option1: 'dynamic-bucket.target-row-num': controls the target row number for one bucket.
  • Option2: 'dynamic-bucket.initial-buckets': controls the number of initialized bucket.
Dynamic Bucket only support single write job. Please do not start multiple jobs to write to the same partition (this can lead to duplicate data). Even if you enable 'write-only' and start a dedicated compaction job, it won’t work.

Normal Dynamic Bucket Mode #

When your updates do not cross partitions (no partitions, or primary keys contain all partition fields), Dynamic Bucket mode uses HASH index to maintain mapping from key to bucket, it requires more memory than fixed bucket mode.

Performance:

  1. Generally speaking, there is no performance loss, but there will be some additional memory consumption, 100 million entries in a partition takes up 1 GB more memory, partitions that are no longer active do not take up memory.
  2. For tables with low update rates, this mode is recommended to significantly improve performance.

Normal Dynamic Bucket Mode supports sort-compact to speed up queries. See Sort Compact.

Cross Partitions Upsert Dynamic Bucket Mode #

When you need cross partition upsert (primary keys not contain all partition fields), Dynamic Bucket mode directly maintains the mapping of keys to partition and bucket, uses local disks, and initializes indexes by reading all existing keys in the table when starting stream write job. Different merge engines have different behaviors:

  1. Deduplicate: Delete data from the old partition and insert new data into the new partition.
  2. PartialUpdate & Aggregation: Insert new data into the old partition.
  3. FirstRow: Ignore new data if there is old value.

Performance: For tables with a large amount of data, there will be a significant loss in performance. Moreover, initialization takes a long time.

If your upsert does not rely on too old data, you can consider configuring index TTL to reduce Index and initialization time:

  • 'cross-partition-upsert.index-ttl': The TTL in rocksdb index and initialization, this can avoid maintaining too many indexes and lead to worse and worse performance.

But please note that this may also cause data duplication.