This is a collection of compression algorithms that are used to compress data of different types. The algorithms are optimized for time-series use-cases; many of them assume that adjacent rows will have "similar" values.
Each compression algorithm the API is divided into two parts: a compressor and a decompression iterator. The compressor is used to compress new data.
<algorithm name>_compressor_alloc- creates the compressor<algorithm_name>_compressor_append_null- appends a null<algorithm_name>_compressor_append_value- appends a non-null value<agorithm_name>_compressor_finish- finalizes the compression and returns the compressed data
Data can be read back out using the decompression iterator. An iterator can operate backwards or forwards. There is no random access. The api is
<algorithm_name>_decompression_iterator_from_datum_<forward|reverse>- create a new DatumIterator in the forward or reverse direction.- a DatumIterator has a function pointer called
try_nextthat returns the nextDecompressResult.
A DecompressResult can either be a decompressed value datum, null, or a done marker to indicate that the iterator is done.
Each decompression algorithm also contains send and recv function to get the external binary representations.
CompressionAlgorithmDefinition is a structure that defines function pointers to get forward and reverse iterators
as well as send and recv functions. The definitions array in compression.c contains a CompressionAlgorithmDefinition
for each compression algorithm.
The simple8b rle algorithm is a building block for many of the compression algorithms.
It compresses a series of uint64 values. It compresses the data by packing the values into the least
amount of bits necessary for the magnitude of the int values, using run-length-encoding for large numbers of repeated values,
A complete description is in the header file. Note that this is a header-only implementation as performance
is paramount here as it is used as a primitive in all the other compression algorithms.
for each integer, it takes the delta-of-deltas with the pervious integer, zigzag encodes this deltadelta, then finally simple8b_rle encodes this zigzagged result. This algorithm performs very well when the magnitude of the delta between adjacent values tends not to vary much, and is optimal for fixed rate-of-change.
gorilla encodes floats using the Facebook gorilla algorithm. It stores the
compressed xors of adjacent values. It is one of the few simple algorithms
that compresses floating point numbers reasonably well.
The dictionary mechanism stores data in two parts: a "dictionary" storing each unique value in the dataset (stored as an array, see below) and simple8b_rle compressed list of indexes into the dictionary, ordered by row. This scheme can store any type of data, but will only be a space improvement if the data set is of relatively low cardinality.
The array "compression" method simply stores the data in an array-like structure and does not actually compress it (though TOAST-based compression can be applied on top). It is the compression mechanism used when no other compression mechanism works. It can store any type of data.
The bool compressor is a simple compression algorithm that stores boolean values using the simple8b_rle algorithm only, without any additional processing. During decompression it decompresses the data and stores it in memory as a bitmap. The row based iterators then walk through the bitmap. The bool compressor differs from the other compressors in that it stores the last non-value as a place holder for the null values. This is done to make vectorization easier.
The uuid compressor is a compression algorithm that aims at storing UUID v7 values compressed as much as possible by taking advantage of the timestamp values being present in the UUID.
The first part of the UUID where the timestamp resides is stored using the delta-delta algorithm. The second part of the UUID is stored without compression, as a sequence of uint64 values.
The algorithm checks the cardinality of the values in the compressed batch and based on the cardinality it decides wether it is worth to recompress the batch using the dictionary compression algorithm. In that case it recompresses and stores the UUIDs as a dictionary.
The external method delegates compression to the column's data type. A type
opts in by providing two SQL functions in its own schema,
<typename>_compress(<type>[]) RETURNS bytea and
<typename>_decompress(bytea) RETURNS <type>[]. It is then selected by
default over dictionary/array for that type. NULLs are stripped into a bitmap
before the compress call and re-inserted after decompression, so these functions
only see non-null values. The returned bytea is stored with TOAST compression
disabled. This lets extension types with internal structure (records, histograms)
compress across rows in ways the generic algorithms cannot.
Chunks will be merged during compression if we specify the compress_chunk_time_interval parameter.
This value will be used to merge chunks adjacent on the time dimension if possible. This allows usage
of smaller chunk intervals which are rolled into bigger compressed chunks.
Compression itself is altered by changing the destination compressed chunk from a newly created one to an already existing chunk which satisfies the necessary requirements (is adjacent to the compressed chunk and chunk interval can be increased not to go over compress chunk time interval).
After compression completes, catalog is updated by dropping the compressed chunk and increasing the chunk interval of the adjacent chunk to include its time dimension slice. Chunk constraints are updated as necessary.
When merging such chunks, due to the nature of sequence number ordering, we will inherently be left with chunks where the sequence numbers are not correctly ordered. In order to mitigate this issue, chunks are recompressed immediately. This has obvious performance implications which might make merging chunks not optimal for certain setups.
We have two functions to determine the columns for timescaledb.compress_segmentby and timescaledb.compress_orderby . These functions can be called
by the UI to give good defaults. They can also be called internally when a hypertable has compression enabled
but no values are provided to specify these options.
This function determines a segment-by column to use. It returns a JSONB with the following top-level keys:
- columns: an array of column names that should be used for segment by. Right now it always returns a single column.
- confidence: a number between 0 and 10 (most confident) indicating how sure we are.
- message: a message that should be shown to the user to evaluate the result.
The intuition is as follows:
we use 3 criterias:
- We want to pick an "important" column for querying. We measure "importance", in terms of how early the column comes in an index (i.e. leading columns are very important, others less so). If there are no indexes, all columns will be considered if statistics are populated
- The column has many rows for the same column value so that the segments will have many rows. We establish that a column will have many values if (i) it is not a dimension and (ii) either statistics tell us so (via
stadistinct> 1) or, if statistics aren't populated, we check whether the column is a generated identity or serial column. - When we have multiple qualifying rows, we select the column where rows are spread most evenly across the distinct values.
Naturally, statistics give us more confidence that the column has enough rows per segment. In this case we break ties by preferring columns from unique indexes. Otherwise, we prefer columns from non-unique indexes (we are less likely to run into a unique column there).
Thus, our preference is based on the whether the column is from a unique or regular index as well as the position of the column in the index. Given these preferences, we think ties happened rarely but will be resolved arbitrarily.
One final point: a number of tables don't have any indexed columns that aren't dimensions or serial columns. In this case, we have medium confidence that an empty segment by is correct.
This function determines which order by columns to use. It returns a JSONB with the following top-level keys:
- clauses: an array of column names and sort order key words that shold be used for order by.
- confidence: a number between 0 and 10 (most confident) indicating how sure we are.
- message: a message that should be shown to the user to evaluate the result.
The order by is built in three steps:
- Use the column order in a unique index (removing the segment_by columns).
- Add any dimension columns
- Add the first attribute of any other index (to establish min-max filters on those columns).
All non-dimension columns are returned without a sort specifier (thus using ASC as default). The dimension columns use DESC.