diff options
| -rw-r--r-- | wiki/git/theory-and-practice/index.md | 41 | ||||
| -rw-r--r-- | wiki/git/theory-and-practice/objects.md | 125 | ||||
| -rw-r--r-- | wiki/git/theory-and-practice/refs-and-names.md | 88 |
3 files changed, 254 insertions, 0 deletions
diff --git a/wiki/git/theory-and-practice/index.md b/wiki/git/theory-and-practice/index.md new file mode 100644 index 0000000..f1e8311 --- /dev/null +++ b/wiki/git/theory-and-practice/index.md @@ -0,0 +1,41 @@ +# Git Internals 101 + +Yeah, yeah, another article about "how Git works". There are tons of these +already. Personally, I'm fond of Sitaram Chamarty's [fantastic series of +articles](http://gitolite.com/master-toc.html) explaining Git from both ends, +and of [Git for Computer +Scientists](http://eagain.net/articles/git-for-computer-scientists/). Maybe +you'd rather read those. + +This page was inspired by very specific, recurring issues I've run into while +helping people use Git. I think Git's "porcelain" layer -- its user interface +-- is terrible, and does a bad job of insulating non-expert users from Git's +internals. While I'd love to fix that (and I do contribute to discussions on +that front, too), we still have the `git(1)` UI right now and people still get +into trouble with it right now. + +Git follows the New Jersey approach laid out in Richard Gabriel's [The Rise of +"Worse is Better"]: given the choice between a simple implementation and a +simple interface, Git chooses the simple implementation almost everywhere. +This internal simplicity can give users the leverage to fix the problems that +its horrible user interface leads them into, so these pages will focus on +explaining the simple parts and giving users the tools to examine them. + +Throughout these articles, I've written "Git does X" a lot. Git is +_incredibly_ configurable; read that as "Git does X _by default_". I'll try to +call out relevant configuration options as I go, where it doesn't interrupt +the flow of knowledge. + +* [Objects](objects) +* [Refs and Names](refs-and-names) + +By the way, if you think you're just going to follow the +[many](http://git-scm.com/documentation) +[excellent](http://www.atlassian.com/git/tutorial) +[git](http://try.github.io/levels/1/challenges/1) +[tutorials](https://www.kernel.org/pub/software/scm/git/docs/gittutorial.html) +out there and that you won't need this knowledge, well, you will. You can +either learn it during a quiet time, when you can think and experiment, or you +can learn it when something's gone wrong, and everyone's shouting at each +other. Git's high-level interface doesn't do much to keep you on the sensible +path, and you will eventually need to fix something.
\ No newline at end of file diff --git a/wiki/git/theory-and-practice/objects.md b/wiki/git/theory-and-practice/objects.md new file mode 100644 index 0000000..985e5dd --- /dev/null +++ b/wiki/git/theory-and-practice/objects.md @@ -0,0 +1,125 @@ +# Objects + +Git's basest level is a storage and naming system for things Git calls +"objects". These objects hold the bulk of the data about files and projects +tracked by Git: file contents, directory trees, commits, and so on. Every +object is identified by a SHA-1 hash, which is derived from its contents. + +SHA-1 hashes are obnoxiously long, so Git allows you to substitue any unique +prefix of a SHA-1 hash, so long as it's at least four characters long. If the +hash `0b43b9e3e64793f5a222a644ed5ab074d8fa1024` is present in your repository, +then Git commands will understand `0b43`, `0b43b9`, and other patterns to all +refer to the same object, so long as no other object has the same SHA-1 +prefix. + +## Blobs + +The contents of every file that's ever been stored in a Git repository are +stored as `blob` objects. These objects are very simple: they contain the file +contents, byte for byte. + +## Trees + +File contents (and trees, and Other Things we'll get to later) are tied +together into a directory structure by `tree` objects. These objects contain a +list of records, with one child per record. Each record contains a permissions +field corresponding to the POSIX permissions mask of the object, a type, a +SHA-1 for another object, and a name. + +A directory containing only files might be represented as the tree + + 100644 blob 511542ad6c97b28d720c697f7535897195de3318 config.md + 100644 blob 801ddd5ae10d6282bbf36ccefdd0b052972aa8e2 integrate.md + 100644 blob 61d28155862607c3d5d049e18c5a6903dba1f85e scratch.md + 100644 blob d7a79c144c22775239600b332bfa120775bab341 survival.md + +while a directory with subdirectories would also have some `tree` children: + + 040000 tree f57ef2457a551b193779e21a50fb380880574f43 12factor + 040000 tree 844697ce99e1ef962657ce7132460ad7a38b7584 authnz + 100644 blob 54795f9b774547d554f5068985bbc6df7b128832 cool-urls-can-change.md + 040000 tree fc3f39eb5d1a655374385870b8be56b202be7dd8 dev + 040000 tree 22cbfb2c1d7b07432ea7706c36b0d6295563c69c devops + 040000 tree 0b3e63b4f32c0c3acfbcf6ba28d54af4c2f0d594 git + 040000 tree 5914fdcbd34e00e23e52ba8e8bdeba0902941d3f java + 040000 tree 346f71a637a4f8933dc754fef02515a8809369c4 mysql + 100644 blob b70520badbb8de6a74b84788a7fefe64a432c56d packaging-ideas.md + 040000 tree 73ed6572345a368d20271ec5a3ffc2464ac8d270 people + +## Commits + +Blobs and trees are sufficient to store arbitrary directory trees in Git, and +you could use them that way, but Git is mostly used as a revision-tracking +system. Revisions and their history are represented by `commit` objects, which contain: + + * The SHA-1 hash of the root `tree` object of the commit, + * Zero or more SHA-1 hashes for parent commits, + * The name and email address of the commit's "author", + * The name and email address of the commit's "committer", + * Timestamps representing when the commit was authored and committed, and + * A commit message. + +Commit objects' parent references form a directed acyclic graph; the subgraph +reachable from a specific commit is that commit's _history_. + +When working with Git's user interface, commit parents are given in a +predictable order determined by the `git checkout` and `git merge` commands. + +## Tags + +Git's revision-tracking system supports "tags", which are stable names for +specific configurations. It also, uniquely, supports a concept called an +"annotated tag", represented by the `tag` object type. These annotated tag +objects contain + + * The type and SHA-1 hash of another object, + * The name and email address of the person who created the tag, + * A timestamp representing the moment the tag was created, and + * A tag message. + +## Anonymity + +There's a general theme to Git's object types: no object knows its own name. +Every object only has a name in the context of some containing object, or in +the context of [Git's refs mechanism](refs-and-names), which I'll get to +shortly. This means that the same `blob` object can be reused for multiple +files (or, more probably, the same file in multiple commits), if they happen +to have the same contents. + +This also applies to tag objects, even though their role is part of a system +for providing stable, meaningful names for commits. + +## Examining objects + +* `git cat-file <type> <sha1>`: decodes the object `<sha1>` and prints its + contents to stdout. This prints the object's contents in their raw form, + which is less than useful for `tree` objects. + +* `git cat-file -p <sha1>`: decodes the object `<sha1>` and pretty-prints it. + This pretty-printing stays close to the underlying disk format; it's most + useful for decoding `tree` objects. + +* `git show <sha1>`: decodes the object `<sha1>` and formats its contents to + stdout. For blobs, this is identical to what `git cat-file blob` would do, + but for trees, commits, and tags, the output is reformated to be more + readable. + +## Storage + +Objects are stored in two places in Git: as "loose objects", and in "pack +files". Newly-created objects are initially loose objects, for ease of +manipulation; transferring objects to another repository or running certain +administrative commands can cause them to be placed in pack files for faster +transfer and for smaller storage. + +Loose objects are stored directly on the filesystem, in the Git repository's +`objects` directory. Git takes a two-character prefix off of each object's +SHA-1 hash, and uses that to pick a subdirectory of `objects` to store the +object in. The remainder of the hash forms the filename. Loose objects are +compressed with zlib, to conserve space, but the resulting directory tree can +still be quite large. + +Packed objects are stored together in packed files, which live in the +repository's `objects/pack` directory. These packed files are both compressed +and delta-encoded, allowing groups of similar objects to be stored very +compactly. diff --git a/wiki/git/theory-and-practice/refs-and-names.md b/wiki/git/theory-and-practice/refs-and-names.md new file mode 100644 index 0000000..94874c9 --- /dev/null +++ b/wiki/git/theory-and-practice/refs-and-names.md @@ -0,0 +1,88 @@ +# Refs and Names + +Git's [object system](objects) stores most of the data for projects tracked in +Git, but only provides SHA-1 hashes. This is basically useless if you want to +make practical use of Git, so Git also has a naming mechanism called "refs" +that provide human-meaningful names for objects. + +There are two kinds of refs: + +* "Normal" refs, which are names that resolve directly to SHA-1 hashes. These + are the vast majority of refs in most repositories. + +* "Symbolic" refs, which are names that resolve to other refs. In most + repositories, only a few of these appear. (Circular references are possible + with symbolic refs. Git will refuse to resolve these.) + +Anywhere you could use a SHA-1, you can use a ref instead. Git interprets them +identically, after resolving the ref down to the SHA-1. + +## Namespaces + +Every operation in Git that uses a name of some sort, including branching +(branch names), tagging (tag names), fetching (remote-tracking branch names), +and pushing (many kinds of name) expand those names to refs, using a namespace +convention. The following namespaces are common: + +* `refs/heads/NAME`: branches. The branch name is the ref name with + `refs/heads/` removed. Names generally point to commits. + +* `refs/remotes/REMOTE/NAME`: "remote-tracking" branches. These are maintained + in tandem by `git remote` and `git fetch`, to cache the state of other + repositories. Names generally point to commits. + +* `refs/tags/NAME`: tags. The tag name is the ref name with `refs/heads/` + removed. Names generally point to commits or tag objects. + +* `refs/stash`: The most recent stash entry, as maintained by `git stash`. + (Other stash entries are maintained by a separate system.) Names generally + point to commits. + +Tools can invent new refs for their own purposes, or manipulate existing refs; +the convention is that tools that use refs (which is, as I said, most of them) +respect the state of the ref as if they'd created that state themselves, +rather than sanity-checking the ref before using it. + +## Special refs + +There are a handful of special refs used by Git commands for their own +operation. These refs do _not_ begin with `refs/`: + +* `HEAD`: the "current" commit for most operations. This is set when checking + out a commit, and many revision-related commands default to `HEAD` if not + given a revision to operate on. `HEAD` can either be a symbolic ref + (pointing to a branch ref) or a normal ref (pointing directly to a commit), + and is very frequently a symbolic ref. + +* `MERGE_HEAD`: during a merge, `MERGE_HEAD` resolves to the commit whose + history is being merged. + +* `ORIG_HEAD`: set by operations that change `HEAD` in potentially destructive + ways by resolving `HEAD` before making the change. + +* `CHERRY_PICK_HEAD` is set during `git cherry-pick` to the commit whose + changes are being copied. + +* `FETCH_HEAD` is set by the forms of `git fetch` that fetch a single ref, and + points to the commit the fetched ref pointed to. + +## Examining and manipulating refs + +The `git show-ref` command will list the refs in namespaces under `refs` in +your repository, printing the SHA-1 hashes they resolve to. Pass `--head` to +also include `HEAD`. + +The following commands can be used to manipulate refs directly: + +* `git update-ref <ref> <sha1>` forcibly sets `<ref>` to the passed `<sha1>`. + +* `git update-ref -d <ref>` deletes a ref. + +* `git symbolic-ref <ref>` prints the target of `<ref>`, if `<ref>` is a + symbolic ref. (It will fail with an error message for normal refs.) + +* `git symbolic-ref <ref> <target>` forcibly makes `<ref>` a symbolic ref + pointing to `<target>`. + +Additionally, you can see what ref a given name resolves to using `git +rev-parse --symbolic-full-name <name>` or `git show-ref <name>`.
\ No newline at end of file |