| 7bd805ff Andy McCall ↶ |
1 |
# Instant clone: a lazy working copy (oxygen#132) |
| 2 |
|
| 3 |
Status: design, 2026-10-25. Nothing here is built yet beyond what it names |
| 4 |
as existing. An experimental `oxy clone --lazy` (Linux) follows if this |
| 5 |
measures well (§9). |
| 6 |
|
| 7 |
## 1. The aim |
| 8 |
|
| 9 |
`oxy clone URL` of any repository, however big, gives a working copy in |
| 10 |
about a second: `ls`, `cd`, editors and builds see every file at once, a |
| 11 |
file's bytes arrive the first time something reads it, and `status`, |
| 12 |
`diff`, `commit` and `push` work at once and cost what the edits cost, not |
| 13 |
what the repository weighs. Git's partial clone and VFS for Git do part of |
| 14 |
this (VFS needs Windows and a special server); Oxygen can do all of it |
| 15 |
with what it already has, because every object is content addressed and |
| 16 |
a state's tree is known without its file contents. |
| 17 |
|
| 18 |
## 2. What exists |
| 19 |
|
| 20 |
- **Shallow clones** (spec §17): `oxy clone --depth 0` takes a checkpoint |
| 21 |
of the bookmark's heads and no history. The `checkpoint` message brings |
| 22 |
the checkpoint, its payload blob and **every file's graph blob** (spec |
| 23 |
§11, payload version 2; `Repo::checkpoint_bundle_for`, |
| 24 |
crates/oxy-core/src/repo/checkpoint.rs). A file graph holds the file's |
| 25 |
current text and its tombstones, so this is the whole state's text and |
| 26 |
more: for VS Code most of the clone's bytes. |
| 27 |
- **Lazy blobs** (spec §13.5, §17): `--max-blob N` leaves opaque blobs |
| 28 |
larger than N to be fetched with `objects` on first read; `status` and |
| 29 |
`commit` never need them. |
| 30 |
- **Sparse filters** (`--sparse PATH`): only those paths are materialized. |
| 31 |
- **flow-casfs** (OxygenFlow, `crates/flow-casfs`): a read-only FUSE mount |
| 32 |
of a state from a local repository. The tree comes from the state |
| 33 |
index; a text file is rendered from its graph on first read or stat and |
| 34 |
cached on disk by graph key; an opaque file is its blob. |
| 35 |
- **Content addressing everywhere**: an object fetched later is checked |
| 36 |
against the hash that named it, so a lazily fetched byte is as |
| 37 |
trustworthy as one cloned eagerly (spec §13). |
| 38 |
|
| 39 |
## 3. The design |
| 40 |
|
| 41 |
`oxy clone --lazy URL DIR`: |
| 42 |
|
| 43 |
1. **Refs and the checkpoint skeleton.** Fetch `refs` (with the bookmark's |
| 44 |
attestation, checked against the policy as any clone does), then the |
| 45 |
checkpoint of its heads **without its graph blobs**: a `checkpoint` |
| 46 |
request with a new `graphs: false` field (additive; an older server |
| 47 |
ignores it and sends them, so the clone is merely not lazy). The |
| 48 |
payload's `files` entries name each graph blob by ObjectRef (graphs |
| 49 |
under 2 KiB are inline and come anyway), and its `tree` gives every |
| 50 |
path, mode and node. Clone writes the repository as a shallow one at |
| 51 |
depth 0 whose missing graph blobs are **lazy objects**, like lazy blobs |
| 52 |
today. |
| 53 |
2. **The mount.** DIR is a FUSE mount (lower layer) of that state, built on |
| 54 |
flow-casfs: the directory tree is answered from the payload's tree at |
| 55 |
once; `stat` of a text file needs its length, so the payload gains |
| 56 |
(payload version 3, writer-optional) each file's rendered length, and |
| 57 |
until a server writes it the length comes with the graph on first |
| 58 |
`stat`. A read of a file whose graph is not stored fetches it with |
| 59 |
`objects` (hash checked), renders it, keeps both, and answers. Opaque |
| 60 |
files fetch their blob the same way. A small daemon (`oxy mount`, |
| 61 |
started by the clone, restarted by any oxy command that finds it gone) |
| 62 |
serves the mount; the repository itself stays in `DIR/.oxy`, which is |
| 63 |
a real directory beside the mount (bind or an overlay entry), so oxy |
| 64 |
never reads through FUSE. |
| 65 |
3. **The writable layer.** Writes go to an upper directory |
| 66 |
(`.oxy/overlay/`): a written file is copied up whole on first write |
| 67 |
(its bytes fetched first if they were never read), a new file is |
| 68 |
created there, a removal is a whiteout, a rename is a whiteout plus a |
| 69 |
copy up. This is the ordinary overlay model, done in the daemon so it |
| 70 |
needs no privileges (kernel overlayfs over FUSE needs a user namespace |
| 71 |
and is not available everywhere). |
| 72 |
4. **Status, diff and commit read the overlay only.** The upper layer is |
| 73 |
exactly the set of paths that may differ from the base state, so |
| 74 |
`status` lists it (comparing each copied-up file's bytes with the |
| 75 |
base rendering, which is stored, since copying up read it) and never |
| 76 |
walks the tree. `diff` and `commit` do the same, and the changes they |
| 77 |
build need only the touched files' graphs, all stored by then. |
| 78 |
`commit` clears the overlay entries it recorded: the mount's base |
| 79 |
state moves to the new heads. |
| 80 |
5. **Fetch, pull, checkout.** A fetch brings changes as today (their |
| 81 |
blobs lazy). Moving the working copy to another state swaps the |
| 82 |
mount's base: the new state's tree is known from the changes and the |
| 83 |
stored checkpoint, the overlay is kept when its paths are unchanged |
| 84 |
between the two states (the `--carry` rule, §10.1), and graphs needed |
| 85 |
to render the new state are lazy again. |
| 86 |
6. **Push.** As today: the pushed changes and the blobs they bring. A |
| 87 |
change built on a lazy base never needs the base's unread files. |
| 88 |
|
| 89 |
## 4. Offline and failures |
| 90 |
|
| 91 |
A file read once is kept, so it reads offline. Reading a file never read |
| 92 |
while the hub cannot be reached fails with `EIO`, and the daemon logs the |
| 93 |
path and the reason; `oxy status` says how many files are not local yet. |
| 94 |
`oxy lazy fetch [PATH..]` brings them in ahead (a directory, the whole |
| 95 |
state, or the files a build reads, as a background job), and `oxy lazy |
| 96 |
materialize` turns the working copy into an ordinary one (unmounted, every |
| 97 |
file written), after which nothing is lazy. A hash that does not match is |
| 98 |
an error for that read, never a wrong byte. |
| 99 |
|
| 100 |
## 5. Without FUSE |
| 101 |
|
| 102 |
Where FUSE is not available (macOS without macFUSE, Windows, a container |
| 103 |
without /dev/fuse), `--lazy` falls back to what exists: a depth-0 clone |
| 104 |
with lazy blobs, and `--sparse` paths when given, saying so. Windows |
| 105 |
ProjFS and macOS File Provider are later backends behind the same daemon |
| 106 |
interface (lookup, read, write-through to the upper layer). |
| 107 |
|
| 108 |
## 6. The protocol |
| 109 |
|
| 110 |
- `checkpoint` gains `graphs: false`: leave out graph blobs (and their |
| 111 |
chunks) the payload names by reference. |
| 112 |
- `objects` already fetches any object by hash, with a blob's chunks; it |
| 113 |
needs batching limits only (as `push.batch-bytes` bounds pushes). |
| 114 |
- Optional later: `render` by graph key (the rendered bytes rather than |
| 115 |
the graph with its tombstones), smaller for long-lived files; a client |
| 116 |
checks it by rendering the graph when it has it, so it is an |
| 117 |
optimization a client may refuse. |
| 118 |
- A host should keep `objects` cheap: it is now on the read path of |
| 119 |
editors and builds. OxygenHub's state cache and bucket stores (Strata) |
| 120 |
already serve objects by hash. |
| 121 |
|
| 122 |
## 7. Trust |
| 123 |
|
| 124 |
Nothing is weaker than an eager clone. The clone checks the bookmark's |
| 125 |
attestation and the checkpoint against its heads (a checkpoint is valid |
| 126 |
only if it reproduces from its heads, spec §11, which a lazy client cannot |
| 127 |
check without the graphs: it trusts the attested heads and checks each |
| 128 |
graph it fetches against the checkpoint's reference). `oxy verify` on a |
| 129 |
lazy clone checks what is stored and says how much is not. |
| 130 |
|
| 131 |
## 8. Not in scope |
| 132 |
|
| 133 |
History (`log`, `blame` beyond the checkpoint) on a lazy clone fetches |
| 134 |
changes on demand later (`oxy deepen` exists); the first version answers |
| 135 |
`log` from the checkpoint's change list as a shallow clone does. |
| 136 |
|
| 137 |
## 9. Measurements |
| 138 |
|
| 2aa1854b Andy McCall ↶ |
139 |
On Earth (4 cores, SSD), 2026-10-25, the VS Code history |
| 140 |
(`~/bench/vscode-full9.oxy`, 166k changes; `~/bench/vscode.git` for Git) |
| 141 |
served over loopback by `oxy serve` and `git daemon`, oxy at main |
| 142 |
(0.5.1 + #137), Git 2.53; one run each, a fresh directory, the server |
| 143 |
warm (its checkpoint of the heads already made). Seconds: |
| 7bd805ff Andy McCall ↶ |
144 |
|
| 2aa1854b Andy McCall ↶ |
145 |
|
| 7bd805ff Andy McCall ↶ |
146 |
|
| 2aa1854b Andy McCall ↶ |
147 |
| clone | 91.4 | 30.5 | 457.9 | 849.0 | |
| 148 |
| `ls -R \| wc -l` | 0.11 | 0.11 | 0.54 | 0.11 | |
| 149 |
| read `src/vs/code/electron-main/main.ts` | 0.01 | 0.01 | 0.01 | 0.01 | |
| 150 |
| append a line, `status` | 0.18 | 0.72 | 8.07 | 8.50 | |
| 151 |
| `diff` | 0.10 | 0.03 | 0.23 | 0.23 | |
| 152 |
| `commit` | 0.12 | 0.05 | 0.98 | 0.92 | |
| 153 |
| disk after | 1.7 GB | 693 MB | 977 MB | 979 MB | |
| 7bd805ff Andy McCall ↶ |
154 |
|
| 2aa1854b Andy McCall ↶ |
155 |
Before #137 the depth-0 clone failed after 11 minutes (the state's binary |
| 156 |
files' blobs came with nothing) and, before that fix's other half, wrote |
| 157 |
the checkpoint object by object: stopped after 38 minutes with 57,498 |
| 158 |
loose objects. The full `oxy clone` (12 GB of history) was not run |
| 159 |
tonight. |
| 7bd805ff Andy McCall ↶ |
160 |
|
| 2aa1854b Andy McCall ↶ |
161 |
What this says: |
| 162 |
|
| 163 |
- A depth-0 clone is five times slower than Git's full clone and fifteen |
| 164 |
times slower than its blobless one. The profile (`OXY_TIMINGS=1`, one |
| 165 |
run of 508 s): about 375 s fetching, because the imported history's |
| 166 |
`main` has 33,835 heads and depth 0 still takes the head changes |
| 167 |
themselves (33,837 changes, with their texts); 12 s decoding the |
| 168 |
checkpoint; 120 s of checkout, 111 s of it rendering 19,545 files on one |
| 169 |
core. So depth 0 should take no change at all (the checkpoint carries |
| 170 |
every change's metadata; spec §17's "a checkpoint plus zero history"), |
| 171 |
and rendering should use every core (oxygen#137). |
| 172 |
- `--max-blob 0` without a mount is slower, not faster: the checkout then |
| 173 |
fetches every binary file's blob one request at a time. Lazy contents |
| 174 |
only pay with a mount that reads them when asked. |
| 175 |
- The first `status` after a clone is 8 s (it builds the manifest); Git's |
| 176 |
is 0.2 s. A lazy clone's `status` from the overlay (§3.4) would not |
| 177 |
depend on the tree at all. |
| 178 |
|
| 179 |
So the prototype's bar: clone in seconds (refs, the checkpoint skeleton |
| 180 |
without graphs, the mount), `ls -R` at once from the payload's tree, a |
| 181 |
first read in tens of milliseconds, `status` in milliseconds; and, |
| 182 |
whatever happens to `--lazy`, the depth-0 clone itself should come down |
| 183 |
to Git's order (streamed checkpoint in parts, rendering on every core). |
| 184 |
|
| 7bd805ff Andy McCall ↶ |
185 |
## 10. The prototype (L2) |
| 186 |
|
| 187 |
1. `checkpoint` with `graphs: false` in oxy-core and oxy-server, and a |
| 188 |
repository flag for lazy graph objects (fetch on load, like lazy blobs). |
| 189 |
2. `oxy mount` (Linux, `fuser`, as flow-casfs) with the upper layer, in a |
| 190 |
new crate `oxy-mount` beside the CLI, behind a cargo feature so builds |
| 191 |
without FUSE are unaffected. |
| 192 |
3. `status`/`diff`/`commit` from the overlay when the working copy is |
| 193 |
mounted. |
| 194 |
4. `oxy clone --lazy`, `oxy lazy fetch`, `oxy lazy materialize`. |
| 195 |
5. Measured as in §9; numbers on oxygen#132. |