Some repositories contain a lot of references (e.g. android at 866k, rails at 31k). The existing packed-refs format takes up a lot of space (e.g. 62M), and does not scale with additional references. Lookup of a single reference requires linearly scanning the file.

Atomic pushes modifying multiple references require copying the entire packed-refs file, which can be a considerable amount of data moved (e.g. 62M in, 62M out) for even small transactions (2 refs modified).

Repositories with many loose references occupy a large number of disk blocks from the local file system, as each reference is its own file storing 41 bytes (and another file for the corresponding reflog). This negatively affects the number of inodes available when a large number of repositories are stored on the same filesystem. Readers can be penalized due to the larger number of syscalls required to traverse and read the $GIT_DIR/refs directory.

A reftable file is a portable binary file format customized for reference storage. References are sorted, enabling linear scans, binary search lookup, and range scans.

Storage in the file is organized into variable sized blocks. Prefix compression is used within a single block to reduce disk space. Block size and alignment are tunable by the writer.

Space used, packed-refs vs. reftable:

Scan (read 866k refs), by reference name lookup (single ref from 866k refs), and by SHA-1 lookup (refs with that SHA-1, from 866k refs):

Space used for 149,932 log entries for 43,061 refs, reflog vs. reftable:

References stored in a reftable are peeled, a record for an annotated (or signed) tag records both the tag object, and the object it refers to. This is analogous to storage in the packed-refs format.

Reference names are an uninterpreted sequence of bytes that must pass git-check-ref-format[1] as a valid reference name.

Each entry must have a unique key; repeated keys are disallowed.

All multi-byte, fixed width fields are in network byte order.

Varint encoding is identical to the ofs-delta encoding method used within pack files.

Decoder works as follows:

Blocks are lexicographically ordered by their first reference.

The reftable format accepts both refs/heads/foo and refs/heads/foo/bar as distinct references.

This property is useful for retaining log records in reftable, but may confuse versions of Git using $GIT_DIR/refs directory tree to maintain references. Users of reftable may choose to continue to reject foo and foo/bar type conflicts to prevent problems for peers.

A reftable file has the following high-level structure:

A log-only file omits the ref_block, ref_index, obj_block and obj_index sections, containing only the file header and log block:

In a log-only file, the first log block immediately follows the file header, without padding to block alignment.

The file’s block size is arbitrarily determined by the writer, and does not have to be a power of 2. The block size must be larger than the longest reference name or log entry used in the repository, as references cannot span blocks.

Powers of two that are friendly to the virtual memory system or filesystem (such as 4k or 8k) are recommended. Larger sizes (64k) can yield better compression, with a possible increased cost incurred by readers during access.

The largest block size is 16777215 bytes (15.99 MiB).

Writers may choose to align blocks at multiples of the block size by including padding filled with NUL bytes at the end of a block to round out to the chosen alignment. When alignment is used, writers must specify the alignment with the file header’s block_size field.

Block alignment is not required by the file format. Unaligned files must set block_size = 0 in the file header, and omit padding. Unaligned files with more than one ref block must include the ref index to support fast lookup. Readers must be able to read both aligned and non-aligned files.

Very small files (e.g. a single ref block) may omit padding and the ref index to reduce total file size.

A 24-byte header appears at the beginning of the file:

Aligned files must specify block_size to configure readers with the expected block alignment. Unaligned files must set block_size = 0.

The min_update_index and max_update_index describe bounds for the update_index field of all log records in this file. When reftables are used in a stack for transactions, these fields can order the files such that the prior file’s max_update_index + 1 is the next file’s min_update_index.

A 28-byte header appears at the beginning of the file:

The header is identical to version_number=1, with the 4-byte hash ID ("sha1" for SHA1 and "s256" for SHA-256) appended to the header.

For maximum backward compatibility, it is recommended to use version 1 when writing SHA1 reftables.

The first ref block shares the same block as the file header, and is 24 bytes smaller than all other blocks in the file. The first block immediately begins after the file header, at position 24.

If the first block is a log block (a log-only file), its block header begins immediately at position 24.

A ref block is written as:

Blocks begin with block_type = 'r' and a 3-byte block_len which encodes the number of bytes in the block up to, but not including the optional padding. This is always less than or equal to the file’s block size. In the first ref block, block_len includes 24 bytes for the file header.

The 2-byte restart_count stores the number of entries in the restart_offset list, which must not be empty. Readers can use restart_count to binary search between restarts before starting a linear scan.

Exactly restart_count 3-byte restart_offset values precede the restart_count. Offsets are relative to the start of the block and refer to the first byte of any ref_record whose name has not been prefix compressed. Entries in the restart_offset list must be sorted, ascending. Readers can start linear scans from any of these records.

A variable number of ref_record fill the middle of the block, describing reference names and values. The format is described below.

As the first ref block shares the first file block with the file header, all restart_offset in the first block are relative to the start of the file (position 0), and include the file header. This forces the first restart_offset to be 28.

The ref index stores the name of the last reference from every ref block in the file, enabling reduced disk seeks for lookups. Any reference can be found by searching the index, identifying the containing block, and searching within that block.

The index may be organized into a multi-level index, where the 1st level index block points to additional ref index blocks (2nd level), which may in turn point to either additional index blocks (e.g. 3rd level) or ref blocks (leaf level). Disk reads required to access a ref go up with higher index levels. Multi-level indexes may be required to ensure no single index block exceeds the file format’s max block size of 16777215 bytes (15.99 MiB). To achieve constant O(1) disk seeks for lookups the index must be a single level, which is permitted to exceed the file’s configured block size, but not the format’s max block size of 15.99 MiB.

If present, the ref index block(s) appears after the last ref block.

If there are at least 4 ref blocks, a ref index block should be written to improve lookup times. Cold reads using the index require 2 disk reads (read index, read block), and binary searching < 4 blocks also requires ⇐ 2 reads. Omitting the index block from smaller files saves space.

If the file is unaligned and contains more than one ref block, the ref index must be written.

Index block format:

The index blocks begin with block_type = 'i' and a 3-byte block_len which encodes the number of bytes in the block, up to but not including the optional padding.

The restart_offset and restart_count fields are identical in format, meaning and usage as in ref blocks.

To reduce the number of reads required for random access in very large files the index block may be larger than other blocks. However, readers must hold the entire index in memory to benefit from this, so it’s a time-space tradeoff in both file size and reader memory.

Increasing the file’s block size decreases the index size. Alternatively a multi-level index may be used, keeping index blocks within the file’s block size, but increasing the number of blocks that need to be accessed.

Object blocks are optional. Writers may choose to omit object blocks, especially if readers will not use the object name to ref mapping.

Object blocks use unique, abbreviated 2-31 byte object name keys, mapping to ref blocks containing references pointing to that object directly, or as the peeled value of an annotated tag. Like ref blocks, object blocks use the file’s standard block size. The abbreviation length is available in the footer as obj_id_len.

To save space in small files, object blocks may be omitted if the ref index is not present, as brute force search will only need to read a few ref blocks. When missing, readers should brute force a linear search of all references to lookup by object name.

An object block is written as:

Fields are identical to ref block. Binary search using the restart table works the same as in reference blocks.

Because object names are abbreviated by writers to the shortest unique abbreviation within the reftable, obj key lengths have a variable length. Their length must be at least 2 bytes. Readers must compare only for common prefix match within an obj block or obj index.

The obj index stores the abbreviation from the last entry for every obj block in the file, enabling reduced disk seeks for all lookups. It is formatted exactly the same as the ref index, but refers to obj blocks.

The obj index should be present if obj blocks are present, as obj blocks should only be written in larger files.

Readers loading the obj index must first read the footer (below) to obtain obj_index_position. If not present, the position will be 0.

Unlike ref and obj blocks, log blocks are always unaligned.

Log blocks are variable in size, and do not match the block_size specified in the file header or footer. Writers should choose an appropriate buffer size to prepare a log block for deflation, such as 2 * block_size.

A log block is written as:

Log blocks look similar to ref blocks, except block_type = 'g'.

The 4-byte block header is followed by the deflated block contents using zlib deflate. The block_len in the header is the inflated size (including 4-byte block header), and should be used by readers to preallocate the inflation output buffer. A log block’s block_len may exceed the file’s block size.

Offsets within the log block (e.g. restart_offset) still include the 4-byte header. Readers may prefer prefixing the inflation output buffer with the 4-byte header.

Within the deflate container, a variable number of log_record describe reference changes. The log record format is described below. See ref block format (above) for a description of restart_offset and restart_count.

Because log blocks have no alignment or padding between blocks, readers must keep track of the bytes consumed by the inflater to know where the next log block begins.

The log index stores the log key (refname \0 reverse_int64(update_index)) for the last log record of every log block in the file, supporting bounded-time lookup.

A log index block must be written if 2 or more log blocks are written to the file. If present, the log index appears after the last log block. There is no padding used to align the log index to block alignment.

Log index format is identical to ref index, except the keys are 9 bytes longer to include '\0' and the 8-byte reverse_int64(update_index). Records use block_position to refer to the start of a log block.

After the last block of the file, a file footer is written. It begins like the file header, but is extended with additional data.

If a section is missing (e.g. ref index) the corresponding position field (e.g. ref_index_position) will be 0.

The size of the footer is 68 bytes for version 1, and 72 bytes for version 2.

Binary search within a block is supported by the restart_offset fields at the end of the block. Readers can binary search through the restart table to locate between which two restart points the sought reference or key should appear.

Each record identified by a restart_offset stores the complete key in the suffix field of the record, making the compare operation during binary search straightforward.

Once a restart point lexicographically before the sought reference has been identified, readers can linearly scan through the following record entries to locate the sought record, terminating if the current record sorts after (and therefore the sought key is not present).

The reftable format assumes the vast majority of references are single object names valued with common prefixes, such as Gerrit Code Review’s refs/changes/ namespace, GitHub’s refs/pulls/ namespace, or many lightweight tags in the refs/tags/ namespace.

Annotated tags storing the peeled object cost an additional object name per reference.

A reftable with very few references (e.g. git.git with 5 heads) is 269 bytes for reftable, vs. 332 bytes for packed-refs. This supports reftable scaling down for transaction logs (below).

For a Gerrit Code Review type repository with many change refs, larger block sizes (64 KiB) and less frequent restart points (every 64) yield better compression due to more references within the block compressing against the prior reference.

Larger block sizes reduce the index size, as the reftable will require fewer blocks to store the same number of references.

Assuming the index block has been loaded into memory, binary searching for any single reference requires exactly 1 disk seek to load the containing block.

Scanning all references and lookup by name (or namespace such as refs/heads/) are the most common activities performed on repositories. Object names are stored directly with references to optimize this use case.

Logs are infrequently accessed, but can be large. Deflating log blocks saves disk space, with some increased penalty at read time.

Logs are stored in an isolated section from refs, reducing the burden on reference readers that want to ignore logs. Further, historical logs can be isolated into log-only files.

Logs are frequently accessed backwards (most recent N records for master to answer master@{4}), so log records are grouped by reference, and sorted descending by update index.

A repository must set its $GIT_DIR/config to configure reftable:

A collection of reftable files are stored in the $GIT_DIR/reftable/ directory. Their names should have a random element, such that each filename is globally unique; this helps avoid spurious failures on Windows, where open files cannot be removed or overwritten. It suggested to use ${min_update_index}-${max_update_index}-${random}.ref as a naming convention.

Log-only files use the .log extension, while ref-only and mixed ref and log files use .ref. extension.

The stack ordering file is $GIT_DIR/reftable/tables.list and lists the current files, one per line, in order, from oldest (base) to newest (most recent):

Readers must read $GIT_DIR/reftable/tables.list to determine which files are relevant right now, and search through the stack in reverse order (last reftable is examined first).

Reftable files not listed in tables.list may be new (and about to be added to the stack by the active writer), or ancient and ready to be pruned.

Older clients should continue to recognize the directory as a git repository so they don’t look for an enclosing repository in parent directories. To this end, a reftable-enabled repository must contain the following dummy files

Readers can obtain a consistent snapshot of the reference space by following:

Although reftables are immutable, mutations are supported by writing a new reftable and atomically appending it to the stack:

During step 4 the new file’s min_update_index and max_update_index are both set to the update_index selected by step 3. All log records for the transaction use the same update_index in their keys. This enables later correlation of which references were updated by the same transaction.

Because a single tables.list.lock file is used to manage locking, the repository is single-threaded for writers. Writers may have to busy-spin (with backoff) around creating tables.list.lock, for up to an acceptable wait period, aborting if the repository is too busy to mutate. Application servers wrapped around repositories (e.g. Gerrit Code Review) can layer their own lock/wait queue to improve fairness to writers.

Deletion of any reference can be explicitly stored by setting the type to 0x0 and omitting the value field of the ref_record. This serves as a tombstone, overriding any assertions about the existence of the reference from earlier files in the stack.

A partial stack of reftables can be compacted by merging references using a straightforward merge join across reftables, selecting the most recent value for output, and omitting deleted references that do not appear in remaining, lower reftables.

A compacted reftable should set its min_update_index to the smallest of the input files' min_update_index, and its max_update_index likewise to the largest input max_update_index.

For sake of illustration, assume the stack currently consists of reftable files (from oldest to newest): A, B, C, and D. The compactor is going to compact B and C, leaving A and D alone.

This strategy permits compactions to proceed independently of updates.

Each reftable (compacted or not) is uniquely identified by its name, so open reftables can be cached by their name.

On windows, and other systems that do not allow deleting or renaming to open files, compaction may succeed, but other readers may prevent obsolete tables from being deleted.

On these platforms, the following strategy can be followed: on closing a reftable stack, reload tables.list, and delete any tables no longer mentioned in tables.list.

Irregular program exit may still leave about unused files. In this case, a cleanup operation should proceed as follows:

bzip2 can significantly shrink a large packed-refs file (e.g. 62 MiB compresses to 23 MiB, 37%). However the bzip format does not support random access to a single reference. Readers must inflate and discard while performing a linear scan.

Breaking packed-refs into chunks (individually compressing each chunk) would reduce the amount of data a reader must inflate, but still leaves the problem of indexing chunks to support readers efficiently locating the correct chunk.

Given the compression achieved by reftable’s encoding, it does not seem necessary to add the complexity of bzip/gzip/zlib.

Michael Haggerty proposed an alternate format to reftable on the Git mailing list. This format uses smaller chunks, without the restart table, and avoids block alignment with padding. Reflog entries immediately follow each ref, and are thus interleaved between refs.

Performance testing indicates reftable is faster for lookups (51% faster, 11.2 usec vs. 5.4 usec), although reftable produces a slightly larger file (+ ~3.2%, 28.3M vs 29.2M):

JGit Ketch proposed RefTree, an encoding of references inside Git tree objects stored as part of the repository’s object database.

The RefTree format adds additional load on the object database storage layer (more loose objects, more objects in packs), and relies heavily on the packer’s delta compression to save space. Namespaces which are flat (e.g. thousands of tags in refs/tags) initially create very large loose objects, and so RefTree does not address the problem of copying many references to modify a handful.

Flat namespaces are not efficiently searchable in RefTree, as tree objects in canonical formatting cannot be binary searched. This fails the need to handle a large number of references in a single namespace, such as GitHub’s refs/pulls, or a project with many tags.

David Turner proposed using LMDB, as LMDB is lightweight (64k of runtime code) and GPL-compatible license.

A downside of LMDB is its reliance on a single C implementation. This makes embedding inside JGit (a popular reimplementation of Git) difficult, and hoisting onto virtual storage (for JGit DFS) virtually impossible.

A common format that can be supported by all major Git implementations (git-core, JGit, libgit2) is strongly preferred.