{"full_name":"ModernRelay/omnigraph","owner":"ModernRelay","name":"omnigraph","shard":"1","url":"https://github.com/ModernRelay/omnigraph","description":"Object-storage native graph database with branching and typed ontology","homepage":"https://omnigraph.dev","topics":["apache-arrow","datafusion","graph-database","knowledge-graph","lakehouse","lance","mcp","rust","s3","versioning","context-graph"],"tag_pages":["apache-arrow","graph-database","knowledge-graph","lakehouse","mcp","rust","s3"],"language":"Rust","license":"MIT","avatar":"https://avatars.githubusercontent.com/u/206314331?v=4","created_at":"2026-04-10T17:50:30Z","pushed_at":"2026-11-16T09:11:08Z","metrics":{"stars":2700,"forks":322,"watchers":25,"issues":298,"prs":1115,"releases":55,"issues_open":134,"issues_closed":164,"prs_open":43,"prs_closed":156,"prs_merged":916,"discussions":0},"growth":{"day":{"stars":213,"forks":11,"base_day":"2026-11-15"}},"history":{"dates":["2026-11-14","2026-11-15","2026-11-16"],"stars":[2302,2487,2700],"forks":[296,311,322],"watchers":[25,24,25],"issues_open":[141,142,134],"prs_open":[40,42,43]},"boards":{"daily":117,"daily_rate":14,"weekly_rate":79,"monthly_rate":172},"same_owner":[],"issues":[{"number":432,"title":"Feature request: table scoping for `optimize` (--table / --exclude-table)","comments":6,"created_at":"2026-08-03T11:29:02Z"},{"number":752,"title":"Attribute the served read floor (~0.2 s server-side on a trivial read) and expose it as usage","comments":4,"created_at":"2026-09-18T11:29:49Z"},{"number":644,"title":"storage: safely reclaim obsolete physical __manifest versions and files","comments":4,"created_at":"2026-09-05T21:15:29Z"},{"number":643,"title":"performance: merges into independent targets serialize behind graph-wide guards","comments":4,"created_at":"2026-09-05T21:14:45Z"},{"number":415,"title":"server: async job API for /branches/merge — 202 + job_id, request-key idempotency","comments":4,"created_at":"2026-08-02T08:27:56Z"},{"number":1233,"title":"bug: retained branch reads use the original path after whole-root relocation (0.14.0)","comments":3,"created_at":"2026-11-02T23:57:37Z"},{"number":485,"title":"Vector ANN is effectively brute-force: indexes are built with 1 IVF partition, and nearest() sets no probe budget","comments":3,"created_at":"2026-08-11T22:08:33Z"},{"number":434,"title":"omnigraph-server: runtime reload of the bearer-token source (SIGHUP or admin endpoint)","comments":3,"created_at":"2026-08-03T16:11:21Z"}],"discussions":[],"latest_release":{"name":"v0.15.0","tag":"v0.15.0","published_at":"2026-11-09T07:58:48Z","body":"Omnigraph 0.15 has five main highlights. Loading a large dataset into a new graph takes less than half the time it took in 0.14. Reads no longer wait for writes to commit. Most branch diffs and merges cost what changed, not how big the tables are or how long the history is. MCP agents get every graph operation, and your Cedar policy alone decides what each agent may run. S3 clusters close three ways in which two servers could write at once.\n\nThis release changes what MCP agents may call, how `@card` minimums and `@unique` are checked, managed organization commands, saved plans and a Rust error type. Read the upgrade actions before you upgrade. If you serve MCP, agents gain every tool that their policy allows, even if you never set an MCP flag. Graph storage does not change, and existing graphs open without a rebuild.\n\nThis page says what each change means for you. The [complete changelog](https://github.com/ModernRelay/omnigraph/blob/v0.15.0/docs/releases/v0.15.0.md) keeps the change notes as they were published.\n\n## Highlights\n\n### CSV bulk load: a new graph in under half the 0.14 time\n\nAn initial import, or a rebuild from an export, now finishes in less than half the time. In local runs of the LDBC SNB SF30 dataset (34 GB of CSV, 31 load statements into a new graph, release builds on an Apple silicon workstation with local disk), the load took 261 seconds on 0.14.0 and 87 to 110 seconds on 0.15. Loads from S3 were not measured. An overwrite `load csv` now works as a pipeline: each parse thread reads its own 8 MiB pieces of the source, checks keys and edge endpoints as rows arrive, and writes data files while the next rows parse. Checks that a new graph cannot fail are skipped, and a refused load still leaves nothing behind. `append` and `merge` loads still read the whole source before they parse it. #1141, #1147, #1158, #1178\n\n### Reads no longer wait for writes\n\nWithin one server process, a read that arrived while a write was committing used to wait for the commit, because the commit held a lock across a storage round trip; on S3 that is a network request. In 0.15 a read never waits for a write in the same process. A write that touches several tables, such as a node type and its edges, also commits them in parallel: two round trips at commit, where it took two for each table, so the next write on the branch starts sooner. The first read after another process commits sends 11 or 12 storage requests instead of 18 to 20. How it works: the graph's current state is an immutable value behind an atomic pointer, and a write swaps the pointer in one step. #1228, #1270, #1230\n\n### Merges after 1,024 commits: 54 storage requests instead of 4,168\n\nReviewing or merging a branch used to cost more the larger the table and the longer its history: a diff read every row of each changed type, and a merge read every data file of each table. Now both read only what changed. Diffing ten updated rows in a type with 100,000 rows took 277 requests and 6.9 MB; it now takes 14 requests and 27 KB, the same as with 4,000 rows. After 1,024 one-row commits, a fast-forward merge sends 54 requests instead of 4,168, and a three-way merge 135 instead of 3,216. These counts come from Omnigraph's simulated in-memory store; on S3 each request is a network round trip. Lance keeps a stable ID for each row and the version that last wrote it, so Omnigraph reads only the rows whose version differs. A diff between two branches that have both moved, and a merge of a table with Blob properties, still compare everything. #1201, #1229,\n#1301\n\n### MCP: every graph operation, decided by your policy\n\nAn agent connected over MCP can now review a branch diff, find a merge base, publish or drop a proposed commit, and run a stored mutation; in 0.14 it needed HTTP for all of these. Only byte streams (Blob payloads and exports) stay HTTP and CLI only. Each tool runs its HTTP route's code under the same Cedar action, so an agent gets the same answers and refusals as an HTTP client","is_prerelease":false},"readme":true,"readme_zh":false,"news":[],"stale":false,"snapshot_at":"2026-11-16T09:23:14Z","history_start":"2026-09-06"}