8000 docs(roadmap): enterprise sources, clusters/sizing, on-prem model providers; add SIZING_AND_SLA.md by xerj-org · Pull Request #2994 · xerj-org/xerj · GitHub
[go: up one dir, main page]

Skip to content

docs(roadmap): enterprise sources, clusters/sizing, on-prem model providers; add SIZING_AND_SLA.md - #2994

Open
xerj-org wants to merge 1 commit into
mainfrom
docs/roadmap-enterprise
Open

xerj-org wants to merge 1 commit into
mainfrom
docs/roadmap-enterprise

Conversation

@xerj-org
@xerj-org xerj-org commented Nov 6, 2026
Copy link
Copy Markdown
Owner

Adds three planned-work sections to ROADMAP.md and a new sizing page, docs/SIZING_AND_SLA.md. Enterprise users running XERJ on their own hardware ask for search over more than files, for a cluster they can size and rely on, and for the model providers they already run.

ROADMAP.md

The review line, section structure and Open defects shortlist are unchanged. The shortlist still lists #1122 and #1170, and both are open on the tracker. The three new sections sit before Beyond 1.0 — themes:

docs/SIZING_AND_SLA.md

  • Says plainly that there is no formal SLA.
  • States the hardware: 32-core x86_64 with 119 GB RAM, shared with other builds. Measured on 2026-11-05 against rc.101 dfbab650d.
  • Single-node sizing table, from 1 GB to 10 GB of raw text.
  • Search, ingest, recovery and idle numbers.
  • Two cluster tables: what works and what does not.
  • A recipe for sizing a deployment today.
  • Notes that the Capacity planning block on landing/docs/operations.html predates these runs.

Checks

Caveats

  • The numbers are single runs on a shared host, and the lab scripts are not in the repo.
  • Status lines for open PRs say "in review". They need updating when those PRs merge.

🤖 Generated with Claude Code

…viders; add SIZING_AND_SLA.md

Motivation: enterprise users running XERJ on their own hardware ask for search
on top of more than files (databases, vector stores, archives), for a cluster
they can size and rely on, and for the model providers they already run
(Cohere among them). The roadmap said nothing about sources beyond files,
carried clustering as one line under "Beyond 1.0", and no page said how to
size a node.

ROADMAP.md, three new sections before "Beyond 1.0 - themes" (structure, the
review line and the Open defects shortlist are unchanged):
- Enterprise data sources beyond files: archives (#2982, merged after the
  rc.101 tag), live Postgres logical-replication sync with status/lag (#2984,
  open), MySQL binlog (planned, no code), rows to facts for XERJ Query-1
  (#2988, open), vector-DB importers for Qdrant/Chroma/pgvector/Pinecone
  (planned, no code).
- Clusters, scaling and sizing: what the 2026-11-05 rc.101 (dfbab65) lab
  runs measured (0 of 1,404,558 acked ids lost after kill -9; election
  170-405 ms; one dead node blocks writes to 2 of 3 shards, 40 of 60 got 503;
  _count 93/104/103 vs 300; _mget 2 of 6; health counts dead nodes;
  17-25% of acked docs 404 after a 4-shard index), the open fixes (#2992
  per-index shard routing 994/6,000 -> 0/6,000 missed GETs; #2993 in-sync
  shrink 26,802-doc replica gap -> 0; #2987 leader no-op commit; #2991 xerj
  index exit 1 when docs are lost, rc.101 dropped 127,138 docs and exited 0),
  and nine ordered deferred items (promotion, PreVote/CheckQuorum, read
  fan-out, honest health, scale-out, ops, memory/ingest, a real multi-process
  CI job plus a measured cluster SLA run, dead-code cleanup).
- Enterprise on-prem integrations: model providers. Supported embedders and
  rerankers, then Cohere as user-requested: Rerank (#2985, on main, not
  released), Embed (#2983, open; native /v2/embed planned), Command/North tool
  use (#2983 tools + #2990 page, open; Query-1 tool and North connector
  planned). Off by default, no live Cohere run yet, no affiliation.
"Distributed clustering maturity" now points at the new section.

docs/SIZING_AND_SLA.md (new): states there is no formal SLA, the hardware
(32-core x86_64, 119 GB, shared host) and date, the single-node sizing table
(1 GB raw: 4 cores / 8-10 GB RAM / 2.5 GB disk; 5 GB: 8 / 16 / 12; 10 GB:
8-16 / 24 / 25), search (5 GB, 8 cores: 196 qps p99 8.4 ms at 1 client,
747 qps p99 20.4 ms at 8; ~110 qps per busy core), ingest (_bulk 8,415
docs/s, p99 11.6 s), recovery (kill -9 with 726 MB WAL: 24.0 s, 200,000 of
200,000 docs), cluster works / does-not-work tables, and how to size today.
It flags that the operations page's capacity-planning block predates these
runs.

Checks: release-notes gate check 1a logic run locally: shortlist #1122 open,
#1170 open. cargo test -p xerj-engine --test docs_capability_lists: 15
passed, 0 failed. No landing page touched (docs/*.md has no site generator),
so no SEO registration.
@cla-bot cla-bot Bot added the cla-signed label Nov 6, 2026

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant

0