Apache Druid stores its index in segment files, which are partitioned by time. In a basic setup, one segment file is created for each time interval, where the time interval is configurable in the segmentGranularity parameter of the granularitySpec. For Druid to operate well under heavy query load, it is important for the segment file size to be within the recommended range of 300MB-700MB. If your segment files are larger than this range, then consider either changing the granularity of the time interval or partitioning your data and tweaking the targetPartitionSize in your partitionsSpec (a good starting point for this parameter is 5 million rows). See the sharding section below and the ‘Partitioning specification’ section of the Batch ingestion documentation for more information.

A segment file’s core data structures

Here we describe the internal structure of segment files, which is essentially columnar: the data for each column is laid out in separate data structures. By storing each column separately, Druid can decrease query latency by scanning only those columns actually needed for a query. There are three basic column types: the timestamp column, dimension columns, and metric columns, as illustrated in the image below:

Druid column types

The timestamp and metric columns are simple: behind the scenes each of these is an array of integer or floating point values compressed with LZ4. Once a query knows which rows it needs to select, it simply decompresses these, pulls out the relevant rows, and applies the desired aggregation operator. As with all columns, if a query doesn’t require a column, then that column’s data is just skipped over.

Dimensions columns are different because they support filter and group-by operations, so each dimension requires the following three data structures:

  1. A dictionary that maps values (which are always treated as strings) to integer IDs,
  2. A list of the column’s values, encoded using the dictionary in 1, and
  3. For each distinct value in the column, a bitmap that indicates which rows contain that value.

Why these three data structures? The dictionary simply maps string values to integer ids so that the values in (2) and (3) can be represented compactly. The bitmaps in (3) — also known as inverted indexes allow for quick filtering operations (specifically, bitmaps are convenient for quickly applying AND and OR operators). Finally, the list of values in (2) is needed for group by and TopN queries. In other words, queries that solely aggregate metrics based on filters do not need to touch the list of dimension values stored in (2).

To get a concrete sense of these data structures, consider the ‘page’ column from the example data above. The three data structures that represent this dimension are illustrated in the diagram below.

  1. 1: Dictionary that encodes column values
  2. {
  3. "Justin Bieber": 0,
  4. "Ke$ha": 1
  5. }
  6. 2: Column data
  7. [0,
  8. 0,
  9. 1,
  10. 1]
  11. 3: Bitmaps - one for each unique value of the column
  12. value="Justin Bieber": [1,1,0,0]
  13. value="Ke$ha": [0,0,1,1]

Note that the bitmap is different from the first two data structures: whereas the first two grow linearly in the size of the data (in the worst case), the size of the bitmap section is the product of data size * column cardinality. Compression will help us here though because we know that for each row in ‘column data’, there will only be a single bitmap that has non-zero entry. This means that high cardinality columns will have extremely sparse, and therefore highly compressible, bitmaps. Druid exploits this using compression algorithms that are specially suited for bitmaps, such as roaring bitmap compression.

Multi-value columns

If a data source makes use of multi-value columns, then the data structures within the segment files look a bit different. Let’s imagine that in the example above, the second row were tagged with both the ‘Ke$ha’ and ‘Justin Bieber’ topics. In this case, the three data structures would now look as follows:

  1. 1: Dictionary that encodes column values
  2. {
  3. "Justin Bieber": 0,
  4. "Ke$ha": 1
  5. }
  6. 2: Column data
  7. [0,
  8. [0,1], <--Row value of multi-value column can have array of values
  9. 1,
  10. 1]
  11. 3: Bitmaps - one for each unique value
  12. value="Justin Bieber": [1,1,0,0]
  13. value="Ke$ha": [0,1,1,1]
  14. ^
  15. |
  16. |
  17. Multi-value column has multiple non-zero entries

Note the changes to the second row in the column data and the Ke$ha bitmap. If a row has more than one value for a column, its entry in the ‘column data’ is an array of values. Additionally, a row with n values in ‘column data’ will have n non-zero valued entries in bitmaps.

SQL Compatible Null Handling

By default, Druid string dimension columns use the values '' and null interchangeably and numeric and metric columns can not represent null at all, instead coercing nulls to 0. However, Druid also provides a SQL compatible null handling mode, which must be enabled at the system level, through druid.generic.useDefaultValueForNull. This setting, when set to false, will allow Druid to at ingestion time create segments whose string columns can distinguish '' from null, and numeric columns which can represent null valued rows instead of 0.

String dimension columns contain no additional column structures in this mode, instead just reserving an additional dictionary entry for the null value. Numeric columns however will be stored in the segment with an additional bitmap whose set bits indicate null valued rows. In addition to slightly increased segment sizes, SQL compatible null handling can incur a performance cost at query time as well, due to the need to check the null bitmap. This performance cost only occurs for columns that actually contain nulls.

Naming Convention

Identifiers for segments are typically constructed using the segment datasource, interval start time (in ISO 8601 format), interval end time (in ISO 8601 format), and a version. If data is additionally sharded beyond a time range, the segment identifier will also contain a partition number.

An example segment identifier may be: datasource_intervalStart_intervalEnd_version_partitionNum

Segment Components

Behind the scenes, a segment is comprised of several files, listed below.

  • version.bin

    4 bytes representing the current segment version as an integer. E.g., for v9 segments, the version is 0x0, 0x0, 0x0, 0x9

  • meta.smoosh

    A file with metadata (filenames and offsets) about the contents of the other smoosh files

  • XXXXX.smoosh

    There are some number of these files, which are concatenated binary data

    The smoosh files represent multiple files “smooshed” together in order to minimize the number of file descriptors that must be open to house the data. They are files of up to 2GB in size (to match the limit of a memory mapped ByteBuffer in Java). The smoosh files house individual files for each of the columns in the data as well as an index.drd file with extra metadata about the segment.

    There is also a special column called __time that refers to the time column of the segment. This will hopefully become less and less special as the code evolves, but for now it’s as special as my Mommy always told me I am.

In the codebase, segments have an internal format version. The current segment format version is v9.

Format of a column

Each column is stored as two parts:

  1. A Jackson-serialized ColumnDescriptor
  2. The rest of the binary for the column

A ColumnDescriptor is essentially an object that allows us to use Jackson’s polymorphic deserialization to add new and interesting methods of serialization with minimal impact to the code. It consists of some metadata about the column (what type is it, is it multi-value, etc.) and then a list of serialization/deserialization logic that can deserialize the rest of the binary.

Compression

Druid compresses blocks of values for string, long, float, and double columns, using LZ4 by default, and bitmaps for string columns and numeric null values are compressed using Roaring. We recommend sticking with these defaults unless experimental verification with your own data and query patterns suggest that non-default options will perform better in your specific case. For example, for bitmap in string columns, the differences between using Roaring and CONCISE are most pronounced for high cardinality columns. In this case, Roaring is substantially faster on filters that match a lot of values, but in some cases CONCISE can have a lower footprint due to the overhead of the Roaring format (but is still slower when lots of values are matched). Currently, compression is configured on at the segment level rather than individual columns, see IndexSpec for more details.

Sharding Data to Create Segments

Sharding

Multiple segments may exist for the same interval of time for the same datasource. These segments form a block for an interval. Depending on the type of shardSpec that is used to shard the data, Druid queries may only complete if a block is complete. That is to say, if a block consists of 3 segments, such as:

sampleData_2011-01-01T02:00:00:00Z_2011-01-01T03:00:00:00Z_v1_0

sampleData_2011-01-01T02:00:00:00Z_2011-01-01T03:00:00:00Z_v1_1

sampleData_2011-01-01T02:00:00:00Z_2011-01-01T03:00:00:00Z_v1_2

All 3 segments must be loaded before a query for the interval 2011-01-01T02:00:00:00Z_2011-01-01T03:00:00:00Z completes.

The exception to this rule is with using linear shard specs. Linear shard specs do not force ‘completeness’ and queries can complete even if shards are not loaded in the system. For example, if your real-time ingestion creates 3 segments that were sharded with linear shard spec, and only two of the segments were loaded in the system, queries would return results only for those 2 segments.

Schema changes

Replacing segments

Druid uniquely identifies segments using the datasource, interval, version, and partition number. The partition number is only visible in the segment id if there are multiple segments created for some granularity of time. For example, if you have hourly segments, but you have more data in an hour than a single segment can hold, you can create multiple segments for the same hour. These segments will share the same datasource, interval, and version, but have linearly increasing partition numbers.

  1. foo_2015-01-01/2015-01-02_v1_0
  2. foo_2015-01-01/2015-01-02_v1_1
  3. foo_2015-01-01/2015-01-02_v1_2

In the example segments above, the dataSource = foo, interval = 2015-01-01/2015-01-02, version = v1, and partitionNum = 0. If at some later point in time, you reindex the data with a new schema, the newly created segments will have a higher version id.

  1. foo_2015-01-01/2015-01-02_v2_0
  2. foo_2015-01-01/2015-01-02_v2_1
  3. foo_2015-01-01/2015-01-02_v2_2

Druid batch indexing (either Hadoop-based or IndexTask-based) guarantees atomic updates on an interval-by-interval basis. In our example, until all v2 segments for 2015-01-01/2015-01-02 are loaded in a Druid cluster, queries exclusively use v1 segments. Once all v2 segments are loaded and queryable, all queries ignore v1 segments and switch to the v2 segments. Shortly afterwards, the v1 segments are unloaded from the cluster.

Note that updates that span multiple segment intervals are only atomic within each interval. They are not atomic across the entire update. For example, you have segments such as the following:

  1. foo_2015-01-01/2015-01-02_v1_0
  2. foo_2015-01-02/2015-01-03_v1_1
  3. foo_2015-01-03/2015-01-04_v1_2

v2 segments will be loaded into the cluster as soon as they are built and replace v1 segments for the period of time the segments overlap. Before v2 segments are completely loaded, your cluster may have a mixture of v1 and v2 segments.

  1. foo_2015-01-01/2015-01-02_v1_0
  2. foo_2015-01-02/2015-01-03_v2_1
  3. foo_2015-01-03/2015-01-04_v1_2

In this case, queries may hit a mixture of v1 and v2 segments.

Different schemas among segments

Druid segments for the same datasource may have different schemas. If a string column (dimension) exists in one segment but not another, queries that involve both segments still work. Queries for the segment missing the dimension will behave as if the dimension has only null values. Similarly, if one segment has a numeric column (metric) but another does not, queries on the segment missing the metric will generally “do the right thing”. Aggregations over this missing metric behave as if the metric were missing.