github.com/2lambda123/git-lfs@v2.5.2+incompatible/docs/extensions.md (about) 1 # Extending LFS 2 3 Teams who use Git LFS often have custom requirements for how the pointer files 4 and blobs should be handled. Some examples of extensions that could be built: 5 6 * Compress large files on clean, uncompress them on smudge/fetch 7 * Encrypt files on clean, decrypt on smudge/fetch 8 * Scan files on clean to make sure they don't contain sensitive information 9 10 The basic extensibility model is that LFS extensions must be registered 11 explicitly, and they will be invoked on clean and smudge to manipulate the 12 contents of the files as needed. On clean, LFS itself ensures that the pointer 13 file is updated with all the information needed to be able to smudge correctly, 14 and the extensions never modify the pointer file directly. 15 16 NOTE: This feature is considered experimental, and included so developers can 17 work on extensions. Exact details of how extensions work are subject to change 18 based on feedback. It is possible for buggy extensions to leave your repository 19 in a bad state, so don't rely on them with a production git repository without 20 extensive testing. 21 22 ## Registration 23 24 To register an LFS extension, it must be added to the Git config. Each 25 extension needs to define: 26 27 * Its unique name. This will be used as part of the key in the pointer file. 28 * The command to run on clean (when files are added to git). 29 * The command to run on smudge (when files are downloaded and checked out). 30 * The priority of the extension, which must be a unique, non-negative integer. 31 32 The sequence `%f` in the clean and smudge commands will be replaced by the 33 filename being processed. 34 35 Here's an example extension registration in the Git config: 36 37 ``` 38 [lfs "extension.foo"] 39 clean = foo clean %f 40 smudge = foo smudge %f 41 priority = 0 42 [lfs "extension.bar"] 43 clean = bar clean %f 44 smudge = bar smudge %f 45 priority = 1 46 ``` 47 48 ## Clean 49 50 When staging a file, Git invokes the LFS clean filter, as described earlier. If 51 no extensions are installed, the LFS clean filter reads bytes from STDIN, 52 calculates the SHA-256 signature, and writes the bytes to a temp file. It then 53 moves the temp file into the appropriate place in .git/lfs/objects and writes a 54 valid pointer file to STDOUT. 55 56 When an extension is installed, LFS will invoke the extension to do additional 57 processing on the bytes before writing them into the temp file. If multiple 58 extensions are installed, they are invoked in the order defined by their 59 priority. LFS will also insert a key in the pointer file for each extension 60 that was invoked, indicating both the order that the extension was invoked and 61 the oid of the file before that extension was invoked. All of that information 62 is required to be able to reliably smudge the file later. Each new line in the 63 pointer file will be of the form: 64 65 `ext-{order}-{name} {hash-method}:{hash-of-input-to-extension}` 66 67 This naming ensures that all extensions are written in both alphabetical and 68 priority order, and also shows the progression of changes to the oid as it is 69 processed by the extensions. 70 71 Here's an example sequence, assuming extensions foo and bar are installed, as 72 shown in the previous section. 73 74 * Git passes the original contents of the file to LFS clean over STDIN. 75 * LFS reads those bytes and calculates the original SHA-256 signature. 76 * LFS streams the bytes to STDIN of `foo clean`, which is expected to write 77 those bytes, modified or not, to its STDOUT. 78 * LFS reads the bytes from STDOUT of `foo clean`, calculates the SHA-256 79 signature, and writes them to STDIN of `bar clean`, which then writes those 80 bytes, modified or not, to its STDOUT. 81 * LFS reads the bytes from STDOUT of `bar clean`, calculates the SHA-256 82 signature, and writes the bytes to a temp file. 83 * When finished, LFS atomically moves the temp file into `.git/lfs/objects`. 84 * LFS generates the pointer file, with some changes: 85 * The oid and size keys are calculated from the final bytes written to LFS 86 local storage. 87 * LFS also writes keys named `ext-0-foo` and `ext-1-bar` into the pointer, along 88 with their respective input oids. 89 90 Here's an example pointer file, for a file processed by extensions foo and bar: 91 92 ``` 93 version https://git-lfs.github.com/spec/v1 94 ext-0-foo sha256:{original hash} 95 ext-1-bar sha256:{hash after foo} 96 oid sha256:{hash after bar} 97 size 123 98 (ending \n) 99 ``` 100 101 Note: as an optimization, if an extension just does a pass-through, its key can 102 be omitted from the pointer file. This will make smudging the file a bit more 103 efficient since that extension can be skipped. LFS can detect a pass-through 104 extension because the input and output oids will be the same. 105 106 This implies that extensions must have no side effects other than writing to 107 their STDOUT. Otherwise LFS has no way to know what extensions modified a file. 108 109 ## Smudge 110 111 When a file is checked out, Git invokes the LFS smudge filter, as described 112 earlier. If no extensions are installed, the LFS smudge filter inspects the 113 first 100 bytes of the bytes off STDIN, and if it is a pointer file, uses the 114 oid to find the correct object in the LFS storage, and writes those bytes to 115 STDOUT so that Git can write them to the working directory. 116 117 If the pointer file indicates that extensions were invoked on that file, then 118 those extensions must be installed in order to smudge. If they are not 119 installed, not found, or unusable for any reason, LFS will fail to smudge the 120 file, and outputs an error indicating which extension is missing. 121 122 Each of the extensions indicated in the pointer file must be invoked in reverse 123 order to undo the changes they made to the contents of the file. After each 124 extension is invoked, LFS will compare the SHA-256 signature of the bytes output 125 by the extension with the oid stored in the pointer file as the original input 126 to that same extension. Those signatures must match, otherwise the extension 127 did not undo its changes correctly. In that case, LFS fails to smudge the file, 128 and outputs an error indicating which extension is failing. 129 130 Here's an example sequence, indicating how LFS will smudge the pointer file 131 shown in the previous section: 132 133 * Git passes the bytes of the pointer file to LFS smudge over STDIN. Note that 134 when using `git lfs checkout`, LFS reads the files directly from disk rather 135 than off STDIN. The rest of the steps are unaffected either way. 136 * LFS reads those bytes and inspects them to see if this is a pointer file. If 137 it was not, the bytes would just be passed through to STDOUT. 138 * Since it is a pointer file, LFS reads the whole file off STDIN, parses it, and 139 determines that extensions foo and bar both processed the file, in that order. 140 * LFS uses the value of the oid key to find the blob in the `.git/lfs/objects` 141 folder, or download from the server as needed. 142 * LFS writes the contents of the blob to STDIN of `bar smudge`, which modifies 143 them as needed and writes them to its STDOUT. 144 * LFS reads the bytes from STDOUT of `bar smudge`, calculates the SHA-256 145 signature, and writes the bytes to STDIN of `foo smudge`, which modifies them 146 as needed and writes to them its STDOUT. 147 * LFS reads the bytes from STDOUT of `foo smudge`, calculates the SHA-256 148 signature, and writes the bytes to its own STDOUT. 149 * At the end, ensure that the hashes calculated on the outputs of foo and bar 150 match their corresponding input hashes from the pointer file. If not, write a 151 descriptive error message indicating which extension failed to undo its 152 changes. 153 * Question: On error, should we overwrite the file in the working directory with 154 the original pointer file? Can this be done reliably? 155 156 ## Handling errors 157 158 If there are errors in the configuration of LFS extensions, such as invalid 159 extension names, duplicate priorities, etc, then any LFS commands that rely on 160 them will abort with a descriptive error message. 161 162 If an extension is unable to perform its task, it can indicate this error by 163 returning a non-zero error code and writing a descriptive error message to its 164 STDERR. The behavior on an error depends on whether we are cleaning or smudging. 165 166 ### Clean 167 168 If an extension fails to clean a file, it will return a non-zero error code and 169 write an error message to its STDERR. Because the file was not cleaned 170 correctly, it can't be added to the index. LFS will ensure that no pointer file 171 is added or updated for failed files. In addition, it will display the error 172 messages for any files that could not be cleaned (and keep those errors in a 173 log), so that the user can diagnose the failure, and then rerun "git add" on 174 those files. 175 176 ### Smudge 177 178 If an extension fails to smudge a file, it will return a non-zero error code and 179 write an error message to its STDERR. Because the file was not smudged 180 correctly, LFS cannot update that file in the working directory. LFS will 181 ensure that the pointer file is written to both the index and working directory. 182 In addition, it will display the error messages for any files that could not be 183 smudged (and keep those errors in a log), so that the user can diagnose the 184 failure and then rerun `git-lfs checkout` to fix up any remaining pointer files.