You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
prometheus/tsdb/docs/refs.md

5.5 KiB

An overview of different Series and Chunk reference types

Used by callers of TSDB

Location Series access Chunk access
Global interfaces SeriesRef (in postings list) chunks.ChunkRef (ChunkReader interface, Meta.Ref)
Head HeadSeriesRef (autoinc) HeadChunkRef (could be head chunk or mmapped chunk. 5/3B split)
blocks BlockSeriesRef (16B aligned) BlockChunkRef (4/4B split)

SeriesRef

Note: we cover the implementations as used in Prometheus. Other projects may use different implementations.

HeadSeriesRef

HeadSeriesRef is simply a 64bit counter that increments when a new series comes in. Due to series churn, the set of actively used HeadSeriesRefs may be well above zero (e.g. 0-10M may not be used, and 10M-11M is active)

Usage:

  • stripeSeries (note: when you don't know a HeadSeriesRef for a series, you can also access it by a hash of the series' labels)
  • WAL
  • HeadChunkRefs include them for addressing head chunks, as those are owned by the memSeries.

Notes:

  1. M-mapped Head chunks, while they use HeadSeriesRefs, don't contain an index and depend on the series listing in memory. Once mmapped, chunks have HeadSeriesRefs inside them, allowing you to recreate the index from reading chunks (Along with WAL which has the labels for those HeadSeriesRefs. It also has all those samples, but by using m-mapped chunks we can save cpu/time and not replay all of WAL on startup)

  2. During querying, HeadSeriesRef are limited to 2^40 (see HeadChunkRef)

  3. The last HeadSeriesRef is always replayed from the WAL and is continued from there.

BlockSeriesRef

Persistent blocks are independent entities and the format/structure is completely different from head block.

In blocks, series are lexicographically ordered by labels and the byte offset in the index file (divided by 16 because they're all aligned on 16 bytes) becomes the BlockSeriesRef.

They are not sequential because index entries may be multiples of 16 bytes. And they don't start from 0 because the byte offset is absolute and includes the magic number, symbols table, etc.

BlockSeriesRef are only 32 bits for now, because 64 bits would slow down the postings lists disk access. (note: this limits the index size to 2^32 * 16 = 64 GB)

See also:

ChunkRef

Chunk references are used to load chunk data during query execution. Note: we cover the implementations as used in Prometheus. Other projects may use different implementations.

HeadChunkRef

A HeadChunkRef is an 8 byte integer that packs together:

  • 5 Bytes for HeadSeriesRef.
  • 3 Bytes for HeadChunkID (uint64) (see below).

There are two implications here:

  • While HeadSeriesRefs can during ingestion go higher, during querying they are limited to 2^40. Querying too high numbers will lead to query failures (but not impact ingestion).
  • ChunkID keeps growing as we enter new chunks until Prometheus restarts. If Prometheus runs too long, we might hit 2^24. (957 years at 1 sample per 15 seconds). If ChunkID=len(mmappedchunks) then it's the head chunk.

BlockChunkRef

A BlockChunkRef is an 8 byte integer. Unlike HeadChunkRef, it is static and independent of factors such as Prometheus restarting.

It packs together:

  • 4 Bytes for chunk file index in the block. This number just increments. Filenames start at 1 but the BlockChunkRef start at 0.
  • 4 Bytes for the byte offset within the file.

Why does HeadChunkRef contain a series reference and BlockChunkRef does not?

The ChunkRef types allow retrieving the chunk data as efficiently as possible.

  • In the Head block the chunks are in the series struct. So we need to reach the series before we can access the chunk from it. Hence we need to pack the HeadSeriesRef to get to the series.
  • In persistent blocks, the chunk files are separated from the index and static. Hence you only need the co-ordinates within the chunks directory to get to the chunk. Hence no need of BlockSeriesRef.

Used internally in TSDB

  • HeadChunkID references a chunk of a memSeries (either an mmappedChunk or headChunk). If a caller has, for whatever reason, an "old" HeadChunkID that refers to a chunk that has been compacted into a block, querying the memSeries for it will not return any data.
  • ChunkDiskMapperRef is an 8 Byte integer. 4 Bytes are used to refer to a chunks file number and 4 bytes serve as byte offset (similar to BlockChunkRef). mmappedChunk provide this value such that callers can load the mmapped chunk from disk.