path: root/Documentation/git-fast-import.txt

git-fast-import(1)
==================

NAME
----
git-fast-import - Backend for fast Git data importers.


SYNOPSIS
--------
frontend | 'git-fast-import' [options]

DESCRIPTION
-----------
This program is usually not what the end user wants to run directly.
Most end users want to use one of the existing frontend programs,
which parses a specific type of foreign source and feeds the contents
stored there to git-fast-import (gfi).

gfi reads a mixed command/data stream from standard input and
writes one or more packfiles directly into the current repository.
When EOF is received on standard input, fast import writes out
updated branch and tag refs, fully updating the current repository
with the newly imported data.

The gfi backend itself can import into an empty repository (one that
has already been initialized by gitlink:git-init[1]) or incrementally
update an existing populated repository.  Whether or not incremental
imports are supported from a particular foreign source depends on
the frontend program in use.


OPTIONS
-------
--date-format=<fmt>::
	Specify the type of dates the frontend will supply to
	gfi within `author`, `committer` and `tagger` commands.
	See ``Date Formats'' below for details about which formats
	are supported, and their syntax.

--force::
	Force updating modified existing branches, even if doing
	so would cause commits to be lost (as the new commit does
	not contain the old commit).

--max-pack-size=<n>::
	Maximum size of each output packfile, expressed in MiB.
	The default is 4096 (4 GiB) as that is the maximum allowed
	packfile size (due to file format limitations). Some
	importers may wish to lower this, such as to ensure the
	resulting packfiles fit on CDs.

--depth=<n>::
	Maximum delta depth, for blob and tree deltification.
	Default is 10.

--active-branches=<n>::
	Maximum number of branches to maintain active at once.
	See ``Memory Utilization'' below for details.  Default is 5.

--export-marks=<file>::
	Dumps the internal marks table to <file> when complete.
	Marks are written one per line as `:markid SHA-1`.
	Frontends can use this file to validate imports after they
	have been completed.

--quiet::
	Disable all non-fatal output, making gfi silent when it
	is successful.	This option disables the output shown by
	\--stats.

--stats::
	Display some basic statistics about the objects gfi has
	created, the packfiles they were stored into, and the
	memory used by gfi during this run.  Showing this output
	is currently the default, but can be disabled with \--quiet.


Performance
-----------
The design of gfi allows it to import large projects in a minimum
amount of memory usage and processing time.  Assuming the frontend
is able to keep up with gfi and feed it a constant stream of data,
import times for projects holding 10+ years of history and containing
100,000+ individual commits are generally completed in just 1-2
hours on quite modest (~$2,000 USD) hardware.

Most bottlenecks appear to be in foreign source data access (the
source just cannot extract revisions fast enough) or disk IO (gfi
writes as fast as the disk will take the data).  Imports will run
faster if the source data is stored on a different drive than the
destination Git repository (due to less IO contention).


Development Cost
----------------
A typical frontend for gfi tends to weigh in at approximately 200
lines of Perl/Python/Ruby code.  Most developers have been able to
create working importers in just a couple of hours, even though it
is their first exposure to gfi, and sometimes even to Git.  This is
an ideal situation, given that most conversion tools are throw-away
(use once, and never look back).


Parallel Operation
------------------
Like `git-push` or `git-fetch`, imports handled by gfi are safe to
run alongside parallel `git repack -a -d` or `git gc` invocations,
or any other Git operation (including `git prune`, as loose objects
are never used by gfi).

gfi does not lock the branch or tag refs it is actively importing.
After the import, during its ref update phase, gfi tests each
existing branch ref to verify the update will be a fast-forward
update (the commit stored in the ref is contained in the new
history of the commit to be written).  If the update is not a
fast-forward update, gfi will skip updating that ref and instead
prints a warning message.  gfi will always attempt to update all
branch refs, and does not stop on the first failure.

Branch updates can be forced with \--force, but its recommended that
this only be used on an otherwise quiet repository.  Using \--force
is not necessary for an initial import into an empty repository.


Technical Discussion
--------------------
gfi tracks a set of branches in memory.  Any branch can be created
or modified at any point during the import process by sending a
`commit` command on the input stream.  This design allows a frontend
program to process an unlimited number of branches simultaneously,
generating commits in the order they are available from the source
data.  It also simplifies the frontend programs considerably.

gfi does not use or alter the current working directory, or any
file within it.  (It does however update the current Git repository,
as referenced by `GIT_DIR`.)  Therefore an import frontend may use
the working directory for its own purposes, such as extracting file
revisions from the foreign source.  This ignorance of the working
directory also allows gfi to run very quickly, as it does not
need to perform any costly file update operations when switching
between branches.

Input Format
------------
With the exception of raw file data (which Git does not interpret)
the gfi input format is text (ASCII) based.  This text based
format simplifies development and debugging of frontend programs,
especially when a higher level language such as Perl, Python or
Ruby is being used.

gfi is very strict about its input.  Where we say SP below we mean
*exactly* one space.  Likewise LF means one (and only one) linefeed.
Supplying additional whitespace characters will cause unexpected
results, such as branch names or file names with leading or trailing
spaces in their name, or early termination of gfi when it encounters
unexpected input.

Date Formats
~~~~~~~~~~~~
The following date formats are supported.  A frontend should select
the format it will use for this import by passing the format name
in the \--date-format=<fmt> command line option.

`raw`::
	This is the Git native format and is `<time> SP <offutc>`.
	It is also gfi's default format, if \--date-format was
	not specified.
+
The time of the event is specified by `<time>` as the number of
seconds since the UNIX epoch (midnight, Jan 1, 1970, UTC) and is
written as an ASCII decimal integer.
+
The local offset is specified by `<offutc>` as a positive or negative
offset from UTC.  For example EST (which is 5 hours behind UTC)
would be expressed in `<tz>` by ``-0500'' while UTC is ``+0000''.
The local offset does not affect `<time>`; it is used only as an
advisement to help formatting routines display the timestamp.
+
If the local offset is not available in the source material, use
``+0000'', or the most common local offset.  For example many
organizations have a CVS repository which has only ever been accessed
by users who are located in the same location and timezone.  In this
case the offset from UTC can be easily assumed.
+
Unlike the `rfc2822` format, this format is very strict.  Any
variation in formatting will cause gfi to reject the value.

`rfc2822`::
	This is the standard email format as described by RFC 2822.
+
An example value is ``Tue Feb 6 11:22:18 2007 -0500''.  The Git
parser is accurate, but a little on the lenient side.  Its the
same parser used by gitlink:git-am[1] when applying patches
received from email.
+
Some malformed strings may be accepted as valid dates.  In some of
these cases Git will still be able to obtain the correct date from
the malformed string.  There are also some types of malformed
strings which Git will parse wrong, and yet consider valid.
Seriously malformed strings will be rejected.
+
Unlike the `raw` format above, the timezone/UTC offset information
contained in an RFC 2822 date string is used to adjust the date
value to UTC prior to storage.  Therefore it is important that
this information be as accurate as possible.
+
If the source material is formatted in RFC 2822 style dates,
the frontend should let gfi handle the parsing and conversion
(rather than attempting to do it itself) as the Git parser has
been well tested in the wild.
+
Frontends should prefer the `raw` format if the source material
is already in UNIX-epoch format, or is easily convertible to
that format, as there is no ambiguity in parsing.

`now`::
	Always use the current time and timezone.  The literal
	`now` must always be supplied for `<when>`.
+
This is a toy format.  The current time and timezone of this system
is always copied into the identity string at the time it is being
created by gfi.  There is no way to specify a different time or
timezone.
+
This particular format is supplied as its short to implement and
may be useful to a process that wants to create a new commit
right now, without needing to use a working directory or
gitlink:git-update-index[1].
+
If separate `author` and `committer` commands are used in a `commit`
the timestamps may not match, as the system clock will be polled
twice (once for each command).  The only way to ensure that both
author and committer identity information has the same timestamp
is to omit `author` (thus copying from `committer`) or to use a
date format other than `now`.

Commands
~~~~~~~~
gfi accepts several commands to update the current repository
and control the current import process.  More detailed discussion
(with examples) of each command follows later.

`commit`::
	Creates a new branch or updates an existing branch by
	creating a new commit and updating the branch to point at
	the newly created commit.

`tag`::
	Creates an annotated tag object from an existing commit or
	branch.  Lightweight tags are not supported by this command,
	as they are not recommended for recording meaningful points
	in time.

`reset`::
	Reset an existing branch (or a new branch) to a specific
	revision.  This command must be used to change a branch to
	a specific revision without making a commit on it.

`blob`::
	Convert raw file data into a blob, for future use in a
	`commit` command.  This command is optional and is not
	needed to perform an import.

`checkpoint`::
	Forces gfi to close the current packfile, generate its
	unique SHA-1 checksum and index, and start a new packfile.
	This command is optional and is not needed to perform
	an import.

`commit`
~~~~~~~~
Create or update a branch with a new commit, recording one logical
change to the project.

....
	'commit' SP <ref> LF
	mark?
	('author' SP <name> SP LT <email> GT SP <when> LF)?
	'committer' SP <name> SP LT <email> GT SP <when> LF
	data
	('from' SP <committish> LF)?
	('merge' SP <committish> LF)?
	(filemodify | filedelete | filedeleteall)*
	LF
....

where `<ref>` is the name of the branch to make the commit on.
Typically branch names are prefixed with `refs/heads/` in
Git, so importing the CVS branch symbol `RELENG-1_0` would use
`refs/heads/RELENG-1_0` for the value of `<ref>`.  The value of
`<ref>` must be a valid refname in Git.  As `LF` is not valid in
a Git refname, no quoting or escaping syntax is supported here.

A `mark` command may optionally appear, requesting gfi to save a
reference to the newly created commit for future use by the frontend
(see below for format).  It is very common for frontends to mark
every commit they create, thereby allowing future branch creation
from any imported commit.

The `data` command following `committer` must supply the commit
message (see below for `data` command syntax).  To import an empty
commit message use a 0 length data.  Commit messages are free-form
and are not interpreted by Git.  Currently they must be encoded in
UTF-8, as gfi does not permit other encodings to be specified.

Zero or more `filemodify`, `filedelete` and `filedeleteall` commands
may be included to update the contents of the branch prior to
creating the commit.  These commands may be supplied in any order.
However it is recommended that a `filedeleteall` command preceed
all `filemodify` commands in the same commit, as `filedeleteall`
wipes the branch clean (see below).

`author`
^^^^^^^^
An `author` command may optionally appear, if the author information
might differ from the committer information.  If `author` is omitted
then gfi will automatically use the committer's information for
the author portion of the commit.  See below for a description of
the fields in `author`, as they are identical to `committer`.

`committer`
^^^^^^^^^^^
The `committer` command indicates who made this commit, and when
they made it.

Here `<name>` is the person's display name (for example
``Com M Itter'') and `<email>` is the person's email address
(``cm@example.com'').  `LT` and `GT` are the literal less-than (\x3c)
and greater-than (\x3e) symbols.  These are required to delimit
the email address from the other fields in the line.  Note that
`<name>` is free-form and may contain any sequence of bytes, except
`LT` and `LF`.  It is typically UTF-8 encoded.

The time of the change is specified by `<when>` using the date format
that was selected by the \--date-format=<fmt> command line option.
See ``Date Formats'' above for the set of supported formats, and
their syntax.

`from`
^^^^^^
Only valid for the first commit made on this branch by this
gfi process.  The `from` command is used to specify the commit
to initialize this branch from.  This revision will be the first
ancestor of the new commit.

Omitting the `from` command in the first commit of a new branch will
cause gfi to create that commit with no ancestor. This tends to be
desired only for the initial commit of a project.  Omitting the
`from` command on existing branches is required, as the current
commit on that branch is automatically assumed to be the first
ancestor of the new commit.

As `LF` is not valid in a Git refname or SHA-1 expression, no
quoting or escaping syntax is supported within `<committish>`.

Here `<committish>` is any of the following:

* The name of an existing branch already in gfi's internal branch
  table.  If gfi doesn't know the name, its treated as a SHA-1
  expression.

* A mark reference, `:<idnum>`, where `<idnum>` is the mark number.
+
The reason gfi uses `:` to denote a mark reference is this character
is not legal in a Git branch name.  The leading `:` makes it easy
to distingush between the mark 42 (`:42`) and the branch 42 (`42`
or `refs/heads/42`), or an abbreviated SHA-1 which happened to
consist only of base-10 digits.
+
Marks must be declared (via `mark`) before they can be used.

* A complete 40 byte or abbreviated commit SHA-1 in hex.

* Any valid Git SHA-1 expression that resolves to a commit.  See
  ``SPECIFYING REVISIONS'' in gitlink:git-rev-parse[1] for details.

The special case of restarting an incremental import from the
current branch value should be written as:
----
	from refs/heads/branch^0
----
The `{caret}0` suffix is necessary as gfi does not permit a branch to
start from itself, and the branch is created in memory before the
`from` command is even read from the input.  Adding `{caret}0` will force
gfi to resolve the commit through Git's revision parsing library,
rather than its internal branch table, thereby loading in the
existing value of the branch.

`merge`
^^^^^^^
Includes one additional ancestor commit, and makes the current
commit a merge commit.  An unlimited number of `merge` commands per
commit are permitted by gfi, thereby establishing an n-way merge.
However Git's other tools never create commits with more than 15
additional ancestors (forming a 16-way merge).  For this reason
it is suggested that frontends do not use more than 15 `merge`
commands per commit.

Here `<committish>` is any of the commit specification expressions
also accepted by `from` (see above).

`filemodify`
^^^^^^^^^^^^
Included in a `commit` command to add a new file or change the
content of an existing file.  This command has two different means
of specifying the content of the file.

External data format::
	The data content for the file was already supplied by a prior
	`blob` command.  The frontend just needs to connect it.
+
....
	'M' SP <mode> SP <dataref> SP <path> LF
....
+
Here `<dataref>` can be either a mark reference (`:<idnum>`)
set by a prior `blob` command, or a full 40-byte SHA-1 of an
existing Git blob object.

Inline data format::
	The data content for the file has not been supplied yet.
	The frontend wants to supply it as part of this modify
	command.
+
....
	'M' SP <mode> SP 'inline' SP <path> LF
	data
....
+
See below for a detailed description of the `data` command.

In both formats `<mode>` is the type of file entry, specified
in octal.  Git only supports the following modes:

* `100644` or `644`: A normal (not-executable) file.  The majority
  of files in most projects use this mode.  If in doubt, this is
  what you want.
* `100755` or `755`: A normal, but executable, file.
* `120000`: A symlink, the content of the file will be the link target.

In both formats `<path>` is the complete path of the file to be added
(if not already existing) or modified (if already existing).

A `<path>` string must use UNIX-style directory seperators (forward
slash `/`), may contain any byte other than `LF`, and must not
start with double quote (`"`).

If an `LF` or double quote must be encoded into `<path>` shell-style
quoting should be used, e.g. `"path/with\n and \" in it"`.

The value of `<path>` must be in canoncial form. That is it must not:

* contain an empty directory component (e.g. `foo//bar` is invalid),
* end with a directory seperator (e.g. `foo/` is invalid),
* start with a directory seperator (e.g. `/foo` is invalid),
* contain the special component `.` or `..` (e.g. `foo/./bar` and
  `foo/../bar` are invalid).

It is recommended that `<path>` always be encoded using UTF-8.

`filedelete`
^^^^^^^^^^^^
Included in a `commit` command to remove a file from the branch.
If the file removal makes its directory empty, the directory will
be automatically removed too.  This cascades up the tree until the
first non-empty directory or the root is reached.

....
	'D' SP <path> LF
....

here `<path>` is the complete path of the file to be removed.
See `filemodify` above for a detailed description of `<path>`.

`filedeleteall`
^^^^^^^^^^^^^^^
Included in a `commit` command to remove all files (and also all
directories) from the branch.  This command resets the internal
branch structure to have no files in it, allowing the frontend
to subsequently add all interesting files from scratch.

....
	'deleteall' LF
....

This command is extremely useful if the frontend does not know
(or does not care to know) what files are currently on the branch,
and therefore cannot generate the proper `filedelete` commands to
update the content.

Issuing a `filedeleteall` followed by the needed `filemodify`
commands to set the correct content will produce the same results
as sending only the needed `filemodify` and `filedelete` commands.
The `filedeleteall` approach may however require gfi to use slightly
more memory per active branch (less than 1 MiB for even most large
projects); so frontends that can easily obtain only the affected
paths for a commit are encouraged to do so.

`mark`
~~~~~~
Arranges for gfi to save a reference to the current object, allowing
the frontend to recall this object at a future point in time, without
knowing its SHA-1.  Here the current object is the object creation
command the `mark` command appears within.  This can be `commit`,
`tag`, and `blob`, but `commit` is the most common usage.

....
	'mark' SP ':' <idnum> LF
....

where `<idnum>` is the number assigned by the frontend to this mark.
The value of `<idnum>` is expressed as an ASCII decimal integer.
The value 0 is reserved and cannot be used as
a mark.  Only values greater than or equal to 1 may be used as marks.

New marks are created automatically.  Existing marks can be moved
to another object simply by reusing the same `<idnum>` in another
`mark` command.

`tag`
~~~~~
Creates an annotated tag referring to a specific commit.  To create
lightweight (non-annotated) tags see the `reset` command below.

....
	'tag' SP <name> LF
	'from' SP <committish> LF
	'tagger' SP <name> SP LT <email> GT SP <when> LF
	data
	LF
....

where `<name>` is the name of the tag to create.

Tag names are automatically prefixed with `refs/tags/` when stored
in Git, so importing the CVS branch symbol `RELENG-1_0-FINAL` would
use just `RELENG-1_0-FINAL` for `<name>`, and gfi will write the
corresponding ref as `refs/tags/RELENG-1_0-FINAL`.

The value of `<name>` must be a valid refname in Git and therefore
may contain forward slashes.  As `LF` is not valid in a Git refname,
no quoting or escaping syntax is supported here.

The `from` command is the same as in the `commit` command; see
above for details.

The `tagger` command uses the same format as `committer` within
`commit`; again see above for details.

The `data` command following `tagger` must supply the annotated tag
message (see below for `data` command syntax).  To import an empty
tag message use a 0 length data.  Tag messages are free-form and are
not interpreted by Git.  Currently they must be encoded in UTF-8,
as gfi does not permit other encodings to be specified.

Signing annotated tags during import from within gfi is not
supported.  Trying to include your own PGP/GPG signature is not
recommended, as the frontend does not (easily) have access to the
complete set of bytes which normally goes into such a signature.
If signing is required, create lightweight tags from within gfi with
`reset`, then create the annotated versions of those tags offline
with the standard gitlink:git-tag[1] process.

`reset`
~~~~~~~
Creates (or recreates) the named branch, optionally starting from
a specific revision.  The reset command allows a frontend to issue
a new `from` command for an existing branch, or to create a new
branch from an existing commit without creating a new commit.

....
	'reset' SP <ref> LF
	('from' SP <committish> LF)?
	LF
....

For a detailed description of `<ref>` and `<committish>` see above
under `commit` and `from`.

The `reset` command can also be used to create lightweight
(non-annotated) tags.  For example:

====
	reset refs/tags/938
	from :938
====

would create the lightweight tag `refs/tags/938` referring to
whatever commit mark `:938` references.

`blob`
~~~~~~
Requests writing one file revision to the packfile.  The revision
is not connected to any commit; this connection must be formed in
a subsequent `commit` command by referencing the blob through an
assigned mark.

....
	'blob' LF
	mark?
	data
....

The mark command is optional here as some frontends have chosen
to generate the Git SHA-1 for the blob on their own, and feed that
directly to `commit`.  This is typically more work than its worth
however, as marks are inexpensive to store and easy to use.

`data`
~~~~~~
Supplies raw data (for use as blob/file content, commit messages, or
annotated tag messages) to gfi.  Data can be supplied using an exact
byte count or delimited with a terminating line.  Real frontends
intended for production-quality conversions should always use the
exact byte count format, as it is more robust and performs better.
The delimited format is intended primarily for testing gfi.

Exact byte count format::
	The frontend must specify the number of bytes of data.
+
....
	'data' SP <count> LF
	<raw> LF
....
+
where `<count>` is the exact number of bytes appearing within
`<raw>`.  The value of `<count>` is expressed as an ASCII decimal
integer.  The `LF` on either side of `<raw>` is not
included in `<count>` and will not be included in the imported data.

Delimited format::
	A delimiter string is used to mark the end of the data.
	gfi will compute the length by searching for the delimiter.
	This format is primarly useful for testing and is not
	recommended for real data.
+
....
	'data' SP '<<' <delim> LF
	<raw> LF
	<delim> LF
....
+
where `<delim>` is the chosen delimiter string.  The string `<delim>`
must not appear on a line by itself within `<raw>`, as otherwise
gfi will think the data ends earlier than it really does.  The `LF`
immediately trailing `<raw>` is part of `<raw>`.  This is one of
the limitations of the delimited format, it is impossible to supply
a data chunk which does not have an LF as its last byte.

`checkpoint`
~~~~~~~~~~~~
Forces gfi to close the current packfile, start a new one, and to
save out all current branch refs, tags and marks.

....
	'checkpoint' LF
	LF
....

Note that gfi automatically switches packfiles when the current
packfile reaches \--max-pack-size, or 4 GiB, whichever limit is
smaller.  During an automatic packfile switch gfi does not update
the branch refs, tags or marks.

As a `checkpoint` can require a significant amount of CPU time and
disk IO (to compute the overall pack SHA-1 checksum, generate the
corresponding index file, and update the refs) it can easily take
several minutes for a single `checkpoint` command to complete.

Frontends may choose to issue checkpoints during extremely large
and long running imports, or when they need to allow another Git
process access to a branch.  However given that a 30 GiB Subversion
repository can be loaded into Git through gfi in about 3 hours,
explicit checkpointing may not be necessary.


Tips and Tricks
---------------
The following tips and tricks have been collected from various
users of gfi, and are offered here as suggestions.

Use One Mark Per Commit
~~~~~~~~~~~~~~~~~~~~~~~
When doing a repository conversion, use a unique mark per commit
(`mark :<n>`) and supply the \--export-marks option on the command
line.  gfi will dump a file which lists every mark and the Git
object SHA-1 that corresponds to it.  If the frontend can tie
the marks back to the source repository, it is easy to verify the
accuracy and completeness of the import by comparing each Git
commit to the corresponding source revision.

Coming from a system such as Perforce or Subversion this should be
quite simple, as the gfi mark can also be the Perforce changeset
number or the Subversion revision number.

Freely Skip Around Branches
~~~~~~~~~~~~~~~~~~~~~~~~~~~
Don't bother trying to optimize the frontend to stick to one branch
at a time during an import.  Although doing so might be slightly
faster for gfi, it tends to increase the complexity of the frontend
code considerably.

The branch LRU builtin to gfi tends to behave very well, and the
cost of activating an inactive branch is so low that bouncing around
between branches has virtually no impact on import performance.

Use Tag Fixup Branches
~~~~~~~~~~~~~~~~~~~~~~
Some other SCM systems let the user create a tag from multiple
files which are not from the same commit/changeset.  Or to create
tags which are a subset of the files available in the repository.

Importing these tags as-is in Git is impossible without making at
least one commit which ``fixes up'' the files to match the content
of the tag.  Use gfi's `reset` command to reset a dummy branch
outside of your normal branch space to the base commit for the tag,
then commit one or more file fixup commits, and finally tag the
dummy branch.

For example since all normal branches are stored under `refs/heads/`
name the tag fixup branch `TAG_FIXUP`.  This way it is impossible for
the fixup branch used by the importer to have namespace conflicts
with real branches imported from the source (the name `TAG_FIXUP`
is not `refs/heads/TAG_FIXUP`).

When committing fixups, consider using `merge` to connect the
commit(s) which are supplying file revisions to the fixup branch.
Doing so will allow tools such as gitlink:git-blame[1] to track
through the real commit history and properly annotate the source
files.

After gfi terminates the frontend will need to do `rm .git/TAG_FIXUP`
to remove the dummy branch.

Import Now, Repack Later
~~~~~~~~~~~~~~~~~~~~~~~~
As soon as gfi completes the Git repository is completely valid
and ready for use.  Typicallly this takes only a very short time,
even for considerably large projects (100,000+ commits).

However repacking the repository is necessary to improve data
locality and access performance.  It can also take hours on extremely
large projects (especially if -f and a large \--window parameter is
used).  Since repacking is safe to run alongside readers and writers,
run the repack in the background and let it finish when it finishes.
There is no reason to wait to explore your new Git project!

If you choose to wait for the repack, don't try to run benchmarks
or performance tests until repacking is completed.  gfi outputs
suboptimal packfiles that are simply never seen in real use
situations.

Repacking Historical Data
~~~~~~~~~~~~~~~~~~~~~~~~~
If you are repacking very old imported data (e.g. older than the
last year), consider expending some extra CPU time and supplying
\--window=50 (or higher) when you run gitlink:git-repack[1].
This will take longer, but will also produce a smaller packfile.
You only need to expend the effort once, and everyone using your
project will benefit from the smaller repository.


Packfile Optimization
---------------------
When packing a blob gfi always attempts to deltify against the last
blob written.  Unless specifically arranged for by the frontend,
this will probably not be a prior version of the same file, so the
generated delta will not be the smallest possible.  The resulting
packfile will be compressed, but will not be optimal.

Frontends which have efficient access to all revisions of a
single file (for example reading an RCS/CVS ,v file) can choose
to supply all revisions of that file as a sequence of consecutive
`blob` commands.  This allows gfi to deltify the different file
revisions against each other, saving space in the final packfile.
Marks can be used to later identify individual file revisions during
a sequence of `commit` commands.

The packfile(s) created by gfi do not encourage good disk access
patterns.  This is caused by gfi writing the data in the order
it is received on standard input, while Git typically organizes
data within packfiles to make the most recent (current tip) data
appear before historical data.  Git also clusters commits together,
speeding up revision traversal through better cache locality.

For this reason it is strongly recommended that users repack the
repository with `git repack -a -d` after gfi completes, allowing
Git to reorganize the packfiles for faster data access.  If blob
deltas are suboptimal (see above) then also adding the `-f` option
to force recomputation of all deltas can significantly reduce the
final packfile size (30-50% smaller can be quite typical).


Memory Utilization
------------------
There are a number of factors which affect how much memory gfi
requires to perform an import.  Like critical sections of core
Git, gfi uses its own memory allocators to ammortize any overheads
associated with malloc.  In practice gfi tends to ammoritize any
malloc overheads to 0, due to its use of large block allocations.

per object
~~~~~~~~~~
gfi maintains an in-memory structure for every object written in
this execution.  On a 32 bit system the structure is 32 bytes,
on a 64 bit system the structure is 40 bytes (due to the larger
pointer sizes).  Objects in the table are not deallocated until
gfi terminates.  Importing 2 million objects on a 32 bit system
will require approximately 64 MiB of memory.

The object table is actually a hashtable keyed on the object name
(the unique SHA-1).  This storage configuration allows gfi to reuse
an existing or already written object and avoid writing duplicates
to the output packfile.  Duplicate blobs are surprisingly common
in an import, typically due to branch merges in the source.

per mark
~~~~~~~~
Marks are stored in a sparse array, using 1 pointer (4 bytes or 8
bytes, depending on pointer size) per mark.  Although the array
is sparse, frontends are still strongly encouraged to use marks
between 1 and n, where n is the total number of marks required for
this import.

per branch
~~~~~~~~~~
Branches are classified as active and inactive.  The memory usage
of the two classes is significantly different.

Inactive branches are stored in a structure which uses 96 or 120
bytes (32 bit or 64 bit systems, respectively), plus the length of
the branch name (typically under 200 bytes), per branch.  gfi will
easily handle as many as 10,000 inactive branches in under 2 MiB
of memory.

Active branches have the same overhead as inactive branches, but
also contain copies of every tree that has been recently modified on
that branch.  If subtree `include` has not been modified since the
branch became active, its contents will not be loaded into memory,
but if subtree `src` has been modified by a commit since the branch
became active, then its contents will be loaded in memory.

As active branches store metadata about the files contained on that
branch, their in-memory storage size can grow to a considerable size
(see below).

gfi automatically moves active branches to inactive status based on
a simple least-recently-used algorithm.  The LRU chain is updated on
each `commit` command.  The maximum number of active branches can be
increased or decreased on the command line with \--active-branches=.

per active tree
~~~~~~~~~~~~~~~
Trees (aka directories) use just 12 bytes of memory on top of the
memory required for their entries (see ``per active file'' below).
The cost of a tree is virtually 0, as its overhead ammortizes out
over the individual file entries.

per active file entry
~~~~~~~~~~~~~~~~~~~~~
Files (and pointers to subtrees) within active trees require 52 or 64
bytes (32/64 bit platforms) per entry.  To conserve space, file and
tree names are pooled in a common string table, allowing the filename
``Makefile'' to use just 16 bytes (after including the string header
overhead) no matter how many times it occurs within the project.

The active branch LRU, when coupled with the filename string pool
and lazy loading of subtrees, allows gfi to efficiently import
projects with 2,000+ branches and 45,114+ files in a very limited
memory footprint (less than 2.7 MiB per active branch).


Author
------
Written by Shawn O. Pearce <spearce@spearce.org>.

Documentation
--------------
Documentation by Shawn O. Pearce <spearce@spearce.org>.

GIT
---
Part of the gitlink:git[7] suite