Timeline

A horizon store asked every peer for the whole history and threw away what fell below C on arrival, so a horizon clone cost a dense clone's bandwidth. draft-fossil-horizon-00 §6 lets the request carry the horizon and the origin restrict what it sends.

The client sends `pragma horizon HASH RULE` after its path-cone pragmas on every clone and pull, never horizon-structural, and counts the replies that say `pragma horizon-served content` (HorizonServed on the clone and sync results, internal and public). It still classifies every reply locally, so a horizon-ignorant peer stays correct (§6.3).

The handler reads the pragma into the request's cone and serves it through the existing keep machinery rather than a parallel path: the pull inventory's keep set drops the blobs below the horizon (via ConeReferences.HorizonBelow) and every cluster; the clone walk walks the tree of a held-set check-in only, counts a parent as sent only inside the held set, and sends every sub-manifest node at its row while sending no cluster. A crossing delta is inflated through the same sourceKept test the path cone uses, charged to the round's batch. req-clusters under a horizon answers nothing, and every reply says horizon-served content.

A horizon store refuses, through refuseClone, a dense clone and a horizon whose held set is not inside its own, both computed over the store with horizon.Place. That is the spec's subset test; zit refuses any other C or rule. A path-only store now also refuses a horizon-only request, which keeps every path.

Interop with zit at horizon-rfc bbd1435674 holds both ways: zit clones and pulls at a horizon from the handler, the client from `zit serve`, and each side's horizon store refuses the other's dense clone. check-in: 352a0d9e0b user: [email protected] tags: feat/horizon-cones, trunk

zit#266 reported that a horizon clone pulling from stock Fossil never learns a check-in Fossil announced only inside a cluster: xfer.c clusters 100 or more unclustered artifacts and then announces only the cluster. §4.5 at 9bf7ebfe0f binds the cluster prohibition to an origin whose replies carry horizon-served, and has the clone request a horizon-ignorant origin's cluster members as transit and classify what arrives. HZ-4.5-1 is restated; implementing it waits for horizon-served, which the horizon pragmas bring. The sync draft on that branch differs from the trunk copy pinned here only by §9's editorial list of extension pragma names. check-in: d5926b325c user: [email protected] tags: feat/horizon-cones, trunk

A content-horizon store against any peer: below-horizon blobs it is sent are recorded as elided rather than stored, it holds and requests no cluster, and a below-horizon blob an operation needs is fetched on demand. check-in: 27b8439777 user: [email protected] tags: feat/horizon-cones, trunk

No peer speaks the horizon pragmas yet, so a content-horizon store receives whatever a dense peer sends (draft-fossil-horizon-00 §6.3). The received-content filter now applies the horizon's classification, read from internal/content rather than re-derived: blobs below the horizon become elided phantoms and are counted in the result's Elided. Blobs no held manifest names yet stay as unreferenced holdings, and retained delta bases stay transit (§4.2).

A horizon store also acts on no cluster (§4.5). Crosslinking a cluster applies its tag but makes no member phantoms. The filter drops the cluster's content, and classification elides any phantom that carries the cluster tag. The client stops sending req-clusters. Path-cone-only stores keep today's cluster behaviour.

Conformance cases run against go-libfossil's handler and fossil(1): clone and pull, deepen, demand fetch through UpdateContext, transit deltas, and cluster rules. check-in: f1f4240f40 user: [email protected] tags: feat/horizon-cones, trunk

Sparse-store classification sized to what it proves rather than to the store: file content is never read, completely linked check-ins answer from the crosslink tables, and a content horizon reads only its held set's trees. A path-cone store of fossil-scm.org classifies in about 2 s at a 65 MiB peak, where v0.14.0 failed with a byte-budget error. check-in: cd1b25e114 user: [email protected] tags: feat/horizon-cones, trunk

Sparse classification read, inflated and delta-expanded every held blob, file content included, under a whole-store byte budget. On a clone of the Fossil SCM repository (67,894 blobs, 8.4 GB expanded) LoadAbsences failed at once with "byte budget exceeded", so no path-cone or content-horizon store of a real repository could classify its phantoms, and pull, verify and compaction failed with it.

Reference facts now come from the tables wherever they provably stand in for raw cards. A blob the crosslinker recorded as no artifact and mlink names as a file version is never read. A completely linked check-in (event row, no orphan row, a primary parent edge) whose primary parent and baseline trees are themselves reproduced contributes its mlink rows as F cards and its plink, cherrypick and tagxref rows as references; by induction over the primary chain those rows give the same (name, file) pairs as its raw cards. Everything else is still read raw: unlinked and orphaned check-ins, roots, check-ins whose parent never arrived, other artifacts, and every check-in once the store has seen a sub-manifest, since trees are not crosslinked. Budgets are charged per round for the selected blobs and their delta ancestors only, and a base's bytes are released once its last delta is applied.

A content horizon now reads only its held set's manifests and the baselines and nodes they reach, on demand, with one visited set and a byte budget scaled to the held set.

On the Fossil SCM clone, cone {src} classifies in about 2 s at a 65 MiB peak heap, reading 8,655 of 67,894 blobs; the inside/outside sets equal the all-raw result, and Outside/Required agree on every artifact. check-in: b28c06506b user: [email protected] tags: feat/horizon-cones, trunk

The content horizon's store classification: a blob is elided when no check-in in the held set names it through its full tree, a kept blob is in both keep sets under a path cone, structure is never elided under a horizon, and a horizon the store cannot place elides nothing, as zit does. check-in: bc54a61b2c user: [email protected] tags: feat/horizon-cones, trunk

# Conflicts: # conformance/horizon_test.go # conformance/requirements/horizon.tsv check-in: 80c8df6a74 user: [email protected] tags: feat/horizon-cones, trunk

A store cone may now carry a content horizon (draft-fossil-horizon-00 §3.2, §4.1-§4.3). Classification computes the held set K from the store's own check-in graph and walks K's full trees -- F cards directly, through B baselines and through S sub-manifest nodes, with one visited set over structural names -- because mlink records a file version only where it changed, and a blob introduced below C and unchanged through C is named by C's tree alone. Nothing about K is stored: every classification recomputes it, so moving C or losing the cone is a reload, never a migration.

Blobs no K check-in names are elided; with path members as well a blob is kept only when it is in both keep sets. Under a horizon nothing structural is elided, by the horizon or by path members, and delta bases of held artifacts and shunned names keep their classes. A C the store does not hold elides nothing, as zit's classification does: eliding everything would leave the pull that brings a deepened C nothing to ask for. A structural horizon is refused by SetStoreCone until it is implemented.

internal/content now imports internal/horizon, so the held-set tests, which commit through internal/manifest, move to an external test package to avoid an import cycle. check-in: 2ea6479ba9 user: [email protected] tags: feat/horizon-cones, trunk

zit's author called horizon-rfc bbd1435674 ready to implement against, and zit#265 settled the leaves rule there against fossil's leaf: a childless check-in is one no P card names, open or closed, and K under the leaves rule is C's descendants plus every descendant of each check-in where a childless check-in's ancestry meets C's. Both rules now yield a set closed under descendants, so leaves and branch tips derive exactly.

HeldSet follows zit's computation at that check-in (src/sparse.zig, ClassifyState.extend): meeting points are the members of C's ancestry a childless check-in's parent walk reaches without passing one, and a childless check-in is skipped only when it descends from C. The tests that encoded the old rule are restated: a merged-but-open branch tip adds nothing, a closed branch's tip does, the fork's descendants on C's side are held, and an unrelated root adds nothing.

The draft copy, the README pin and the table follow: HZ-3.1-1, HZ-5.2-1 (the seed applies to every entry of K) and HZ-6.2-11 (announce only in reply to a request that pulls) are restated, and HZ-3.1-4 (closure under descendants) is new, with a case checking that no child of a held check-in falls outside K under either rule. check-in: a68f89e414 user: [email protected] tags: feat/horizon-cones, trunk

Both slices defined the descendants/leaves rule: cone.HorizonRule for the serialized member and horizon.Rule for the walk. One concept gets one type; the walk now takes the value the cone carries. check-in: 6d8ccec5da user: [email protected] tags: feat/horizon-cones, trunk

The held set K of draft-fossil-horizon-00 §3.1 under both rules, bounded and computed from the check-in graph the repository holds. Conflicts in the horizon suite and table resolved by keeping both slices' cases and rows. check-in: 7a5ae7cee7 user: [email protected] tags: feat/horizon-cones, trunk

draft-fossil-horizon-00 Section 3.1 defines K as a function of C, the rule and the check-in graph, recomputed by whoever needs it and stored by no one. internal/horizon.HeldSet computes it from the plink and leaf tables: one bounded plink load, then iterative breadth-first walks.

Merge parents are followed both ways, as zit's walk does (src/sparse.zig, ClassifyState.extend and walkParents). Under the leaves rule each open leaf's ancestry is held up to where it meets C's ancestry, inclusive, so every equally near common ancestor the walk reaches is held. "Open leaf" is fossil's leaf table minus active closed tags rather than zit's "no held child", which keeps a below-C branch merged into a descendant held back to its fork, so any two open leaves keep their merge base in K. C that is not a held check-in is ErrNotHeld; phantom parents are never entered.

Section 8 requires the walk to carry visited-node bounds: the load and every walk share the 1<<20 bound of the other graph walks and fail with ErrBound, and cancellation is honoured. The conformance suite gains cases for HZ-3.1-1 and HZ-3.1-3, both PARTIAL until the horizon member and the content horizon exist. check-in: 75ded1b0e0 user: [email protected] tags: feat/horizon-cones, trunk

The horizon member of a cone and its version-3 serialization, byte-identical to zit's (draft-fossil-horizon-00 §3.3). check-in: 6da46813f9 user: [email protected] tags: feat/horizon-cones, trunk

A horizon names one check-in by full hash, a held-set rule and a depth. Version 3 carries it between the count and the flagged members, a distinct version so a reader predating horizons refuses the cone instead of reading the lead-in as path members. Forms without a horizon are unchanged, so their digests hold. A horizon-only cone has no path members and is path-dense. Parse re-serializes and compares, so an upper-case hash or any other non-canonical input is refused.

HZ-3.3-1 and HZ-3.3-2 are now covered. check-in: 5f637a3b62 user: [email protected] tags: feat/horizon-cones, trunk

Moves the horizon copy from 86d4b67372 to e6ee276c89, the check-in the sparse-cone copy is already pinned at. Every change between the two is in §6: §6.1 lets any request carry pragma horizon-cursor, §6.2 adds the SHOULDs to keep the classification and the structural graph per store prefix, and §6.2.6 defines the cursor. Five new rows read them into the table (HZ-6.2-9 to HZ-6.2-13) and HZ-6.1-1 names the cursor; the table is now 50 requirements, none implemented. check-in: be103fe35a user: [email protected] tags: feat/horizon-cones, trunk

chore(release): v0.14.0 changelog leaf check-in: 6d5357235a user: [email protected] tags: main, trunk, v0.14.0

feat: sub-manifest trees and sparse cones (draft-fossil-submanifests-00, draft-fossil-sparse-cone-00) check-in: 7244496a09 user: [email protected] tags: main, trunk

StartServer kills the server's process group and reaps the leader, but fossil server forks a child per connection, and SIGKILL reaches that child asynchronously. On a loaded machine the child could still finish a write after Wait returned, and t.TempDir's RemoveAll failed with "directory not empty" (TestCloneFromRealFossilUncompressedReply, load average ~20). This is the race #352 narrowed; cleanup now also waits, for at most five seconds, until the group has no member left. leaf check-in: f1b6155c8f user: [email protected] tags: feat/submanifest-reader, main, trunk

Five fixes the three-tool round trip and the fossil-scm.org comparison surfaced: an uncovered clone is refused with the push card first, repo status reports the checkout's changes, repo clone remembers last-sync-url, repo verify reports damage, transit phantoms and elision, and adding a directory adds the files beneath it. check-in: b3c5ad7bca user: [email protected] tags: feat/submanifest-reader, main, trunk

TestScenarioClusterConvergence calls content.GenerateClusters once the simulation is over, to check the API on a repository of 500 blobs. Buggify was still armed, so at hostile seeds 11 and 23 a write-path site corrupted a cluster artifact, blob.Store's round-trip check refused it, and the test failed on the refusal. That refusal is the safeguard working. The call now runs with injection suspended, as the invariant checks and fixtures do. check-in: 0be3dddeae user: [email protected] tags: feat/submanifest-reader, main, trunk

Repo.Verify expanded every non-phantom blob on its own with content.Verify. Each expansion re-applies the blob's whole delta chain, so the cost grows with the square of chain length. On a clone of fossil-scm.org's repository (67,894 blobs, chains averaging 532 deltas and reaching 2,546) repo verify took 49 minutes of CPU on both main and this branch, against six and a half minutes for fossil test-integrity.

content.ExpandAll already walks the delta forest from its roots and applies each delta once, binding every result to its name; internal/verify uses it. Verify now goes through it, with the same blob selection (size >= 0), and names the failing rid in its error. The same clone verifies in 19 seconds. check-in: 6f269e32a9 user: [email protected] tags: feat/submanifest-reader, main, trunk

The simulator arms Buggify when it is built, and the scenarios seed their fixtures after that. Two setups dropped blob.Store's error, so a store the injection failed left an empty UUID, and TestScenarioAuthRestrictedCaps panicked in blob.Exists's precondition when it asserted on it. The seed sweep hit this at adversarial seeds 5 and 8.

seedTx suspends injection for the fixture transaction and fails the test if it does not commit. Every unchecked setup transaction in the package now goes through it. check-in: c796b51895 user: [email protected] tags: feat/submanifest-reader, main, trunk

Manage stored a named directory, "." included, as a tracked file, and the next commit failed reading it. A named directory now expands through the checkout's bounded, cone-aware walk to the files beneath it, skipping paths with a dot component as fossil's add does without --dotfiles; a file named explicitly is kept as given. An uncovered path is refused as such before its absence on disk is reported. check-in: 9e28f473cc user: [email protected] tags: feat/submanifest-reader, main, trunk

repo verify hashed the held blobs and printed a fixed line. It now runs the full verifier the library uses for rebuild, prints each issue, and reports transit phantoms apart from elided artifacts, as sparse-cone 5.2 requires of a verification operation: a store complete within its cone verifies healthy and shows its elided count. check-in: 168062ca33 user: [email protected] tags: feat/submanifest-reader, main, trunk

go-libfossil read last-sync-url as the default peer for repo cone store, widening and required-content recovery, but nothing wrote it, so only a repository fossil had cloned had one. repo clone now stores its source as fossil clone does, with any user name and without the password. check-in: caf59be3b2 user: [email protected] tags: feat/submanifest-reader, main, trunk

Run from inside a checkout with the default --dir ".", repo status reported every tracked file MISSING: filepath.WalkDir's root entry is named ".", and the hidden-directory rule skipped it, and with it the whole tree. The command also compared the disk with tip rather than the checked-out version, so a checkout of an older version read as edited.

The command now opens the checkout and prints Checkout.Status, which classifies changes against the checkout's version as fossil's changes command does, cone included, with fossil's labels. Untracked files are no longer listed; fossil status leaves those to fossil extras. check-in: ea23bc8a41 user: [email protected] tags: feat/submanifest-reader, main, trunk

A sparse store that does not cover a clone (sparse-cone 5.3) answered with the error card alone. fossil 2.28's clone client reads a first round without a push card as an empty, finished clone: it exited 0 and left a repository with no project code and no artifacts.

The refusal now goes through refuseClone, the path page_xfer's refusal of an unauthorized clone already took: the reply is reset, the push card naming the repository goes ahead of the error, and no later card is read. fossil then aborts with the refusal and leaves no file. go-libfossil's client, given the codes, re-asks once and stops at the second refusal, as fossil's does. The origin-refusal case now runs fossil clone against the refusing store. check-in: fb6c1d71a1 user: [email protected] tags: feat/submanifest-reader, main, trunk

The safety checks ran while simio's Buggify was armed, so the read-path sites in blob decompression and content expansion truncated or flipped the bytes each check expanded. Every seed at the adversarial and hostile levels then reported a blob-integrity violation that was never persisted; CI runs the normal level only, so nothing caught it.

simio.SuspendBuggify switches injection off and returns a resume that restores the previous state with its random stream intact, so a check mid-run does not change the run's later choices. Every exported check in dst/invariants.go suspends for its duration. A guard test arms Buggify over a clean repository and fails without the change.

Also records in the CHANGELOG that ExtractOpts and RepoOpenCmd are no longer comparable. check-in: dfdf391349 user: [email protected] tags: feat/submanifest-reader, main, trunk

zit #262 settled the pull-inventory cost in the draft: a dense inventory can cost a requester the out-of-cone blobs of check-ins announced in the same inventory, since those names stay transit phantoms until a held manifest classifies them, and the filtering of 6.3 elides them on arrival. That is what the requester here does, so SC-6.2-4 closes. Its case now checks both halves: a blob announced with its check-in is never held, and one announced after it is never requested. check-in: 07d79d60eb user: [email protected] tags: feat/submanifest-reader, main, trunk

The draft now settles zit #257 to #260, so the four rows that waited on it close against code:

- Compaction and local filtering keep every manifest and sub-manifest node, interiors included (5.4, 6.3). RetainProvenance previously let an outside interior whose children were all held leaves go; the draft now requires it held, since it is the range proof that reaches the leaf. - A server answering a cone request acknowledges the reply with pragma cone-served in its leading pragma run (6.3). This server applies a request's whole cone or refuses it, and a refusal resets the reply, so the pragma is never sent for a refused or dense reply. SyncResult and CloneResult report ConeServed beside Elided, now public too. - An empty cone is refused (3.2); a new row and case pin it, locally and over the wire. - The per-round bound covers the tree walk only (6.2). A new row tests that a requester does not act on a dense inventory. It stays partial: an out-of-cone blob announced in the same reply as its check-in cannot be classified yet and is fetched, then filtered; zit #262 asks the draft to account for that cost. check-in: b5a64230b3 user: [email protected] tags: feat/submanifest-reader, main, trunk

A sparse clone sends the spine, in-cone nodes and in-cone blobs of draft 6.2; a held blob no F card names is none of these, so it is not sent, though zit's keep set holds it. check-in: 4529a3f71c user: [email protected] tags: feat/submanifest-reader, main, trunk

The per-round walk serves a cone clone, and nothing but tests reached the prepared index or the service that built it, so both go with their successor index. The whole-store reference loader stays for the pull inventory, compaction, verification and crosslink, and keeps its budget tests. The shun and private gates the prepared tests covered are now checked against the walk. check-in: ef6efa9f1e user: [email protected] tags: feat/submanifest-reader, main, trunk

A sparse walk now sends one clone-cap cookie per round naming the row the next round starts at, with a part-walked check-in's progress. A request that returns to that check-in without echoing its progress comes from a client that drops it; the round steps past the row instead of resending the same batch forever, and the next pull fetches what the walk owed by name. A gimme in a cone request is answered densely, since the requester names it (6.4). The SC-6.2-1 note records the pull inventory's whole-store read, which matches zit's. check-in: 6766dd6cba user: [email protected] tags: feat/submanifest-reader, main, trunk

Revert wrote a restored file with vfile's bits, which a pending merge may have changed, and left them in vfile. Fossil's revert writes the file with the reverted-to version's permission and records it in vfile.isexe and islink, so reverting a file whose only change is a merged executable bit takes the bit back off. versionPaths now carries each file's permission. check-in: da5d6546bb user: [email protected] tags: feat/submanifest-reader, main, trunk

Apply now does what fossil's stash_apply does when a stashed file changed since the stash: it merges the stashed change (pivot: the stash baseline) onto the file on disk with merge_3way, writing fossil's conflict markers and NAME-baseline/-original/-merge on conflict, an empty file for binary content, and reporting each conflict. Previously the stashed bytes were restored verbatim, losing the current change.

Save reverts a file with a pending merge to the checked-out version's executable and symlink bits, and records them in vfile, as fossil's revert does; the stash keeps the merged bits and the pop restores them. checkout.VersionPaths becomes VersionPerms to supply those bits.

The sparse stash-record-only conformance case now expects fossil's result: a file the merge took unchanged keeps the checked-out content after a stash round trip. check-in: 04e3ea8301 user: [email protected] tags: feat/submanifest-reader, main, trunk

Committing deltifies each manifest against its successor, so the first check-in's chain puts later manifests on the wire before their rows. A sparse round skipped those rows as already sent and never walked them, leaving their in-cone files to a later pull. The row is now walked without being sent again. Walk-only clones in the new cases hold no transit phantoms. check-in: 6791486470 user: [email protected] tags: feat/submanifest-reader, main, trunk

A clone round now reads and walks only its batch. What remains of the MUST is a cone-scoped pull, which still classifies the whole held store to filter its igot inventory and gimme answers. The walk also carries its request context explicitly. check-in: cbc0f9a21e user: [email protected] tags: feat/submanifest-reader, main, trunk

A cone-scoped clone no longer classifies the whole held store before its first round. Each round reads only the rows it reaches: structural artifacts go out at their row, and a check-in's row walks its tree, descending only nodes whose key range meets the cone and that its primary parent's tree does not carry over the same range, then sends the in-cone nodes and blobs that walk reaches. Content and sub-manifest rows wait for those walks and are charged as excluded rows are, so a round's work tracks its batch. A check-in whose walk outgrows a round resumes in the next one from progress kept in the clone-cap cookie. Delta bases the round did not reach are served in full form. The walk keeps the depth and visited-node bounds per request.

A round spent wholly inside an excluded run sends nothing yet advances its cursor; a coned clone now continues on that progress instead of reporting that the server stopped sending.

HandleEnd reports the artifacts a cone request classified. check-in: 2909f2a676 user: [email protected] tags: feat/submanifest-reader, main, trunk