Apache Kudu Schema Design

Depending on the type of the column, Kudu columns can be created with the following encoding types.

Plain Encoding

Data is stored in its natural format. For example, int32 values are stored as fixed-size 32-bit little-endian integers.

Bitshuffle Encoding

A block of values is rearranged to store the most significant bit of every value, followed by the second most significant bit of every value, and so on. Finally, the result is LZ4 compressed. Bitshuffle encoding is a good choice for columns that have many repeated values, or values that change by small amounts when sorted by primary key. The bitshuffle project has a good overview of performance and use cases.

Run Length Encoding

Runs (consecutive repeated values) are compressed in a column by storing only the value and the count. Run length encoding is effective for columns with many consecutive repeated values when sorted by primary key.

Dictionary Encoding

A dictionary of unique values is built, and each column value is encoded as its corresponding index in the dictionary. Dictionary encoding is effective for columns with low cardinality. If the column values of a given row set are unable to be compressed because the number of unique values is too high, Kudu will transparently fall back to plain encoding for that row set. This is evaluated during flush.

Prefix Encoding

Common prefixes are compressed in consecutive column values. Prefix encoding can be effective for values that share common prefixes, or the first column of the primary key, since rows are sorted by primary key within tablets.

Each column in a Kudu table can be created with an encoding, based on the type of the column. Columns use plain encoding by default.

Kudu allows per-column compression using the LZ4, Snappy, or zlib compression codecs. By default, columns are stored uncompressed. Consider using compression if reducing storage space is more important than raw scan performance.

Every data set will compress differently, but in general LZ4 is the most efficient codec, while zlib will compress to the smallest data sizes. Bitshuffle-encoded columns are automatically compressed using LZ4, so it is not recommended to apply additional compression on top of this encoding.

As with many traditional relational databases, Kudu’s primary key is a clustered index. All rows within a tablet are kept in primary key sorted order. Kudu scans which specify equality or range constraints on the primary key will automatically skip rows which can not satisfy the predicate. This allows individual rows to be efficiently found by specifying equality constraints on the primary key columns.

Range partitioning distributes rows using a totally-ordered range partition key. Each partition is assigned a contiguous segment of the range partition keyspace. The key must be comprised of a subset of the primary key columns. If the range partition columns match the primary key columns, then the range partition key of a row will equal its primary key. In range partitioned tables without hash partitioning, each range partition will correspond to exactly one tablet.

The initial set of range partitions is specified during table creation as a set of partition bounds and split rows. For each bound, a range partition will be created in the table. Each split will divide a range partition in two. If no partition bounds are specified, then the table will default to a single partition covering the entire key space (unbounded below and above). Range partitions must always be non-overlapping, and split rows must fall within a range partition.

Hash partitioning distributes rows by hash value into one of many buckets. In single-level hash partitioned tables, each bucket will correspond to exactly one tablet. The number of buckets is set during table creation. Typically the primary key columns are used as the columns to hash, but as with range partitioning, any subset of the primary key columns can be used.

Hash partitioning is an effective strategy when ordered access to the table is not needed. Hash partitioning is effective for spreading writes randomly among tablets, which helps mitigate hot-spotting and uneven tablet sizes.

Kudu allows a table to combine multiple levels of partitioning on a single table. Zero or more hash partition levels can be combined with an optional range partition level. The only additional constraint on multilevel partitioning beyond the constraints of the individual partition types, is that multiple levels of hash partitions must not hash the same columns.

When used correctly, multilevel partitioning can retain the benefits of the individual partitioning types, while reducing the downsides of each. The total number of tablets in a multilevel partitioned table is the product of the number of partitions in each level.

Kudu scans will automatically skip scanning entire partitions when it can be determined that the partition can be entirely filtered by the scan predicates. To prune hash partitions, the scan must include equality predicates on every hashed column. To prune range partitions, the scan must include equality or range predicates on the range partitioned columns. Scans on multilevel partitioned tables can take advantage of partition pruning on any of the levels independently.

To illustrate the factors and tradeoffs associated with designing a partitioning strategy for a table, we will walk through some different partitioning scenarios. Consider the following table schema for storing machine metrics data (using SQL syntax and date-formatted timestamps for clarity):

CREATE TABLE metrics (
    host STRING NOT NULL,
    metric STRING NOT NULL,
    time INT64 NOT NULL,
    value DOUBLE NOT NULL,
    PRIMARY KEY (host, metric, time),
);

Range Partitioning

A natural way to partition the metrics table is to range partition on the time column. Let’s assume that we want to have a partition per year, and the table will hold data for 2014, 2015, and 2016. There are at least two ways that the table could be partitioned: with unbounded range partitions, or with bounded range partitions.

The image above shows the two ways the metrics table can be range partitioned on the time column. In the first example (in blue), the default range partition bounds are used, with splits at 2015-01-01 and 2016-01-01. This results in three tablets: the first containing values before 2015, the second containing values in the year 2015, and the third containing values after 2016. The second example (in green) uses a range partition bound of [(2014-01-01), (2017-01-01)], and splits at 2015-01-01 and 2016-01-01. The second example could have equivalently been expressed through range partition bounds of [(2014-01-01), (2015-01-01)], [(2015-01-01), (2016-01-01)], and [(2016-01-01), (2017-01-01)], with no splits. The first example has unbounded lower and upper range partitions, while the second example includes bounds.

Each of the range partition examples above allows time-bounded scans to prune partitions falling outside of the scan’s time bound. This can greatly improve performance when there are many partitions. When writing, both examples suffer from potential hot-spotting issues. Because metrics tend to always be written at the current time, most writes will go into a single range partition.

The second example is more flexible, because it allows range partitions for future years to be added to the table. In the first example, all writes for times after 2016-01-01 will fall into the last partition, so the partition may eventually become too large for a single tablet server to handle.

Hash Partitioning

Another way of partitioning the metrics table is to hash partition on the host and metric columns.

In the example above, the metrics table is hash partitioned on the host and metric columns into four buckets. Unlike the range partitioning example earlier, this partitioning strategy will spread writes over all tablets in the table evenly, which helps overall write throughput. Scans over a specific host and metric can take advantage of partition pruning by specifying equality predicates, reducing the number of scanned tablets to one. One issue to be careful of with a pure hash partitioning strategy, is that tablets could grow indefinitely as more and more data is inserted into the table. Eventually tablets will become too big for an individual tablet server to hold.

Hash and Range Partitioning

The previous examples showed how the metrics table could be range partitioned on the time column, or hash partitioned on the host and metric columns. These strategies have associated strength and weaknesses:

Partitioning Strategies

Strategy	Writes	Reads	Tablet Growth
range(time)	✗ - all writes go to latest partition	✓ - time-bounded scans can be pruned	✓ - new tablets can be added for future time periods
hash(host, metric)	✓ - writes are spread evenly among tablets	✓ - scans on specific hosts and metrics can be pruned	✗ - tablets could grow too large

Hash partitioning is good at maximizing write throughput, while range partitioning avoids issues of unbounded tablet growth. Both strategies can take advantage of partition pruning to optimize scans in different scenarios. Using multilevel partitioning, it is possible to combine the two strategies in order to gain the benefits of both, while minimizing the drawbacks of each.

In the example above, range partitioning on the time column is combined with hash partitioning on the host and metric columns. This strategy can be thought of as having two dimensions of partitioning: one for the hash level and one for the range level. Writes into this table at the current time will be parallelized up to the number of hash buckets, in this case 4. Reads can take advantage of time bound and specific host and metric predicates to prune partitions. New range partitions can be added, which results in creating 4 additional tablets (as if a new column were added to the diagram).

Hash and Hash Partitioning

Kudu can support any number of hash partitioning levels in the same table, as long as the levels have no hashed columns in common.

In the example above, the table is hash partitioned on host into 4 buckets, and hash partitioned on metric into 3 buckets, resulting in 12 tablets. Although writes will tend to be spread among all tablets when using this strategy, it is slightly more prone to hot-spotting than when hash partitioning over multiple independent columns, since all values for an individual host or metric will always belong to a single tablet. Scans can take advantage of equality predicates on the host and metric columns separately to prune partitions.

Multiple levels of hash partitioning can also be combined with range partitioning, which logically adds another dimension of partitioning.