1 # Adding Custom Transfer Agents to LFS
5 Git LFS supports multiple ways to transfer (upload and download) files. In the
6 core client, the basic way to do this is via a one-off HTTP request via the URL
7 returned from the LFS API for a given object. The core client also supports
8 extensions to allow resuming of downloads (via `Range` headers) and uploads (via
9 the [tus.io](http://tus.io) protocol).
11 Some people might want to be able to transfer content in other ways, however.
12 To enable this, git-lfs allows configuring Custom Transfers, which are
13 simply processes which must adhere to the protocol defined later in this
14 document. git-lfs will invoke the process at the start of all transfers,
15 and will communicate with the process via stdin/stdout for each transfer.
17 ## Custom Transfer Type Selection
19 In the LFS API request, the client includes a list of transfer types it
20 supports. When replying, the API server will pick one of these and make any
21 necessary adjustments to the returned object actions, in case the the picked
22 transfer type needs custom details about how to do each transfer.
24 ## Using a Custom Transfer Type without the API server
26 In some cases the transfer agent can figure out by itself how and where
27 the transfers should be made, without having to query the API server.
28 In this case it's possible to use the custom transfer agent directly,
29 without querying the server, by using the following config option:
31 * `lfs.standalonetransferagent`, `lfs.<url>.standalonetransferagent`
33 Specifies a custom transfer agent to be used if the API server URL matches as
34 in `git config --get-urlmatch lfs.standalonetransferagent <apiurl>`.
35 `git-lfs` will not contact the API server. It instead sets stage 2 transfer
36 actions to `null`. `lfs.<url>.standalonetransferagent` can be used to
37 configure a custom transfer agent for individual remotes.
38 `lfs.standalonetransferagent` unconditionally configures a custom transfer
39 agent for all remotes. The custom transfer agent must be specified in
40 a `lfs.customtransfer.<name>` settings group.
42 ## Defining a Custom Transfer Type
44 A custom transfer process is defined under a settings group called
45 `lfs.customtransfer.<name>`, where `<name>` is an identifier (see
46 [Naming](#naming) below).
48 * `lfs.customtransfer.<name>.path`
50 `path` should point to the process you wish to invoke. This will be invoked at
51 the start of all transfers (possibly many times, see the `concurrent` option
52 below) and the protocol over stdin/stdout is defined below in the
53 [Protocol](#protocol) section.
55 * `lfs.customtransfer.<name>.args`
57 If the custom transfer process requires any arguments, these can be provided
58 here. Typically you would only need this if your process was multi-purpose or
59 particularly flexible, most of the time you won't need it.
61 * `lfs.customtransfer.<name>.concurrent`
63 If true (the default), git-lfs will invoke the custom transfer process
64 multiple times in parallel, according to `lfs.concurrenttransfers`, splitting
65 the transfer workload between the processes.
67 If you would prefer that only one instance of the transfer process is invoked,
68 maybe because you want to do your own parallelism internally (e.g. slicing
69 files into parts), set this to false.
71 * `lfs.customtransfer.<name>.direction`
73 Specifies which direction the custom transfer process supports, either
74 `download`, `upload`, or `both`. The default if unspecified is `both`.
78 Each custom transfer must have a name which is unique to the underlying
79 mechanism, and the client and the server must agree on that name. The client
80 will advertise this name to the server as a supported transfer approach, and if
81 the server supports it, it will return relevant object action links. Because
82 these may be very different from standard HTTP URLs it's important that the
83 client and server agree on the name.
85 For example, let's say I've implemented a custom transfer process which uses
86 NFS. I could call this transfer type `nfs` - although it's not specific to my
87 configuration exactly, it is specific to the way NFS works, and the server will
88 need to give me different URLs. Assuming I define my transfer like this, and the
89 server supports it, I might start getting object action links back like
90 `nfs://<host>/path/to/object`
94 The git-lfs client communicates with the custom transfer process via the stdin
95 and stdout streams. No file content is communicated on these streams, only
96 request / response metadata. The metadata exchanged is always in JSON format.
97 External files will be referenced when actual content is exchanged.
99 ### Line Delimited JSON
101 Because multiple JSON messages will be exchanged on the same stream it's useful
102 to delimit them explicitly rather than have the parser find the closing `}` in
103 an arbitrary stream, therefore each JSON structure will be sent and received on
104 a **single line** as per [Line Delimited
105 JSON](https://en.wikipedia.org/wiki/JSON_Streaming#Line_delimited_JSON_2).
107 In other words when git-lfs sends a JSON message to the custom transfer it will
108 be on a single line, with a line feed at the end. The transfer process must
109 respond the same way by writing a JSON structure back to stdout with a single
110 line feed at the end (and flush the output).
114 The protocol consists of 3 stages:
116 #### Stage 1: Intiation
118 Immediately after invoking a custom transfer process, git-lfs sends initiation
119 data to the process over stdin. This tells the process useful information about
122 The message will look like this:
125 { "event": "init", "operation": "download", "remote": "origin", "concurrent": true, "concurrenttransfers": 3 }
128 * `event`: Always `init` to identify this message
129 * `operation`: will be `upload` or `download` depending on transfer direction
130 * `remote`: The Git remote. It can be a remote name like `origin` or an URL
131 like `ssh://git.example.com//path/to/repo`. A standalone transfer agent can
132 use it to determine the location of remote files.
133 * `concurrent`: reflects the value of `lfs.customtransfer.<name>.concurrent`, in
134 case the process needs to know
135 * `concurrenttransfers`: reflects the value of `lfs.concurrenttransfers`, for if
136 the transfer process wants to implement its own concurrency and wants to
137 respect this setting.
139 The transfer process should use the information it needs from the intiation
140 structure, and also perform any one-off setup tasks it needs to do. It should
141 then respond on stdout with a simple empty confirmation structure, as follows:
147 Or if there was an error:
150 { "error": { "code": 32, "message": "Some init failure message" } }
153 #### Stage 2: 0..N Transfers
155 After the initiation exchange, git-lfs will send any number of transfer
156 requests to the stdin of the transfer process, in a serial sequence. Once a
157 transfer request is sent to the process, it awaits a completion response before
158 sending the next request.
162 For uploads the request sent from git-lfs to the transfer process will look
166 { "event": "upload", "oid": "bf3e3e2af9366a3b704ae0c31de5afa64193ebabffde2091936ad2e7510bc03a", "size": 346232, "path": "/path/to/file.png", "action": { "href": "nfs://server/path", "header": { "key": "value" } } }
169 * `event`: Always `upload` to identify this message
170 * `oid`: the identifier of the LFS object
171 * `size`: the size of the LFS object
172 * `path`: the file which the transfer process should read the upload data from
173 * `action`: the `upload` action copied from the response from the batch API.
174 This contains `href` and `header` contents, which are named per HTTP
175 conventions, but can be interpreted however the custom transfer agent wishes
176 (this is an NFS example, but it doesn't even have to be an URL). Generally,
177 `href` will give the primary connection details, with `header` containing any
178 miscellaneous information needed. `action` is `null` for standalone transfer
181 The transfer process should post one or more [progress messages](#progress) and
182 then a final completion message as follows:
185 { "event": "complete", "oid": "bf3e3e2af9366a3b704ae0c31de5afa64193ebabffde2091936ad2e7510bc03a" }
188 * `event`: Always `complete` to identify this message
189 * `oid`: the identifier of the LFS object
191 Or if there was an error in the transfer:
194 { "event": "complete", "oid": "bf3e3e2af9366a3b704ae0c31de5afa64193ebabffde2091936ad2e7510bc03a", "error": { "code": 2, "message": "Explain what happened to this transfer" } }
197 * `event`: Always `complete` to identify this message
198 * `oid`: the identifier of the LFS object
199 * `error`: Should contain a `code` and `message` explaining the error
203 For downloads the request sent from git-lfs to the transfer process will look
207 { "event": "download", "oid": "22ab5f63670800cc7be06dbed816012b0dc411e774754c7579467d2536a9cf3e", "size": 21245, "action": { "href": "nfs://server/path", "header": { "key": "value" } } }
210 * `event`: Always `download` to identify this message
211 * `oid`: the identifier of the LFS object
212 * `size`: the size of the LFS object
213 * `action`: the `download` action copied from the response from the batch API.
214 This contains `href` and `header` contents, which are named per HTTP
215 conventions, but can be interpreted however the custom transfer agent wishes
216 (this is an NFS example, but it doesn't even have to be an URL). Generally,
217 `href` will give the primary connection details, with `header` containing any
218 miscellaneous information needed. `action` is `null` for standalone transfer
221 Note there is no file path included in the download request; the transfer
222 process should create a file itself and return the path in the final response
223 after completion (see below).
225 The transfer process should post one or more [progress messages](#progress) and
226 then a final completion message as follows:
229 { "event": "complete", "oid": "22ab5f63670800cc7be06dbed816012b0dc411e774754c7579467d2536a9cf3e", "path": "/path/to/file.png" }
232 * `event`: Always `complete` to identify this message
233 * `oid`: the identifier of the LFS object
234 * `path`: the path to a file containing the downloaded data, which the transfer
235 process relinquishes control of to git-lfs. git-lfs will move the file into
238 Or, if there was a failure transferring this item:
241 { "event": "complete", "oid": "22ab5f63670800cc7be06dbed816012b0dc411e774754c7579467d2536a9cf3e", "error": { "code": 2, "message": "Explain what happened to this transfer" } }
244 * `event`: Always `complete` to identify this message
245 * `oid`: the identifier of the LFS object
246 * `error`: Should contain a `code` and `message` explaining the error
248 Errors for a single transfer request should not terminate the process. The error
249 should be returned in the response structure instead.
251 The custom transfer adapter does not need to check the SHA of the file content
252 it has downloaded, git-lfs will do that before moving the final content into
257 In order to support progress reporting while data is uploading / downloading,
258 the transfer process should post messages to stdout as follows before sending
259 the final completion message:
262 { "event": "progress", "oid": "22ab5f63670800cc7be06dbed816012b0dc411e774754c7579467d2536a9cf3e", "bytesSoFar": 1234, "bytesSinceLast": 64 }
265 * `event`: Always `progress` to identify this message
266 * `oid`: the identifier of the LFS object
267 * `bytesSoFar`: the total number of bytes transferred so far
268 * `bytesSinceLast`: the number of bytes transferred since the last progress
271 The transfer process should post these messages such that the last one sent
272 has `bytesSoFar` equal to the file size on success.
274 #### Stage 3: Finish & Cleanup
276 When all transfers have been processed, git-lfs will send the following message
277 to the stdin of the transfer process:
280 { "event": "terminate" }
283 On receiving this message the transfer process should clean up and terminate.
284 No response is expected.
288 Any unexpected fatal errors in the transfer process (not errors specific to a
289 transfer request) should set the exit code to non-zero and print information to
290 stderr. Otherwise the exit code should be 0 even if some transfers failed.
292 ## A Note On Verify Actions
294 You may have noticed that that only the `upload` and `download` actions are
295 passed to the custom transfer agent for processing, what about the `verify`
296 action, if the API returns one?
298 Custom transfer agents do not handle the verification process, only the
299 upload and download of content. The verify link is typically used to notify
300 a system *other* than the actual content store after an upload was completed,
301 therefore it makes more sense for that to be handled via the normal API process.