Skip to content
Crosswalker
🚧 Early alpha — building the foundation. See the roadmap →

Challenge 16: Tier 3 stack reconsideration — alternatives to Apache AGE (archived)

Created Updated

Why this exists

Section titled “Why this exists”

Challenge 10 picked Apache AGE on PostgreSQL as the Tier 3 graph engine. Challenge 11’s three deliverables verified empirical claims about AGE and surfaced concerns:

  • Sponsor health: Bitnine Co. (Korean parent of Bitnine Global, AGE’s primary commercial sponsor) was acquired in December 2024 and renamed SKAI Worldwide in January 2025, pivoting toward AI advertising/content production while still maintaining graph DB products. The fact that the primary commercial sponsor pivoted away from databases is a meaningful long-term-stability concern.
  • Release cadence: ~one major release per supported Postgres line per year; PG 18 support PR was open from October 2025 with slow movement
  • Postgres ABI risk: AGE was directly hit by the PostgreSQL 17.1 ABI break in November 2024 (per Crunchy Data’s blog); extension users had to rebuild against the new minor version
  • Single-vendor anchored: while Apache 2.0 license, governance is single-sponsor without Apache Foundation top-level project status

The user’s reaction (in the TL;DR direction log): “AGE is just slow in development it seemed — so I’m not sure about AGE for [Tier 3 default]. Maybe lots of better options for that which we’ve researched.”

This challenge does that reconsideration explicitly.

What to investigate

Section titled “What to investigate”

1. Apache Jena Fuseki as Tier 3 primary

Section titled “1. Apache Jena Fuseki as Tier 3 primary”

Drop AGE entirely; expose the canonical Crosswalker SSSOM data via Jena’s SPARQL endpoint.

Required:

  • Performance: SPARQL query benchmarks on representative SSSOM workload (tens of thousands of mappings; multi-hop traversals)
  • OWL/RDFS inference: Jena natively supports RDFS and a subset of OWL; what’s the practical inference capability for SSSOM/SKOS workloads?
  • Operational story: JVM dependency; container size; memory footprint; multi-user concurrency
  • Federation: can multiple Jena instances federate via SERVICE clauses for cross-vault queries?
  • Pros: Apache Foundation governance (much stronger than AGE’s single-sponsor); mature; supports RDFS/OWL inference; canonical RDF-native fit for SSSOM/SKOS/STRM
  • Cons: JVM (large container, slow startup); SPARQL-only (no SQL surface for tabular consumers)

2. DuckDB-on-server + DuckPGQ as Tier 3 primary

Section titled “2. DuckDB-on-server + DuckPGQ as Tier 3 primary”

Use the same DuckDB engine across Tier 2 (browser) and Tier 3 (server), with DuckPGQ extension providing graph queries via SQL/PGQ.

Required:

  • DuckPGQ stability: it’s a “community extension” since DuckDB v1.0; the October 2025 DuckDB blog post on graph queries used it; CWI labels it “research project” still under development. Is it stable enough for Tier 3 production?
  • Engine unification value: same engine, same SQL dialect, same storage format across both tiers — operational simplicity is real. What’s the cost?
  • Multi-user concurrency: DuckDB historically targets analytical workloads; how does it handle 10–100 concurrent users querying a server-mode instance?
  • Persistence: DuckDB-on-server uses regular files; how does this compare to Postgres durability guarantees?
  • Pros: One engine across tiers; MIT license; DuckDB Foundation governance; weekly releases; SQL/PGQ standard (vs proprietary Cypher dialect)
  • Cons: DuckPGQ not yet WASM-loadable in DuckDB-WASM (so Tier 2 can’t use SQL/PGQ today); not historically positioned for multi-user server workloads

3. TerminusDB as Tier 3 primary (revisit from Ch 11a)

Section titled “3. TerminusDB as Tier 3 primary (revisit from Ch 11a)”

Ch 11 deliverable A recommended TerminusDB as default Tier 3. Ch 11 deliverables B and C recommended AGE+Jena with TerminusDB as optional vault-mirror. The user’s §5.C decision leaned toward AGE+Jena, but if AGE is being dropped, TerminusDB-as-primary becomes more attractive.

Required:

  • Re-evaluate TerminusDB as Tier 3 primary with AGE removed from consideration
  • Operational complexity: SWI-Prolog + Rust storage; Docker-only; server only (no embedded)
  • Performance ceiling: tens of millions of triples comfortably; billions with sufficient RAM
  • DFRNT stewardship since 2025: small Stockholm-based commercial sponsor; v12 shipped December 2025; on-disk format unchanged from v11
  • Pros: native Git-style branch/diff/merge — matches files-canonical ethos; closed-world RDF + JSON-LD; WOQL Datalog query + GraphQL surface; Apache 2.0
  • Cons: SWI-Prolog stack is unusual (operational learning curve); single-vendor stewardship; no embedded mode

4. HelixDB as Tier 3 primary

Section titled “4. HelixDB as Tier 3 primary”

HelixDB — native Rust graph + vector database on LMDB, strongly-typed compiled query language (HelixQL), built-in MCP support for LLM agents.

Required:

  • License: AGPL — restrictive. Acceptable for self-hosted Tier 3? Implications for redistribution?
  • Y Combinator funding: positive signal but young project
  • Vector + graph + MCP integration: collapses three layers Crosswalker would otherwise stitch
  • Performance: vendor-claimed billions-of-queries; need independent validation
  • Pros: AI-agent-ready (MCP); vector-native; Rust = good operational story
  • Cons: AGPL is restrictive; very young project; HelixQL is proprietary query language

5. Other candidates already on the radar

Section titled “5. Other candidates already on the radar”
EngineLicenseWhy might it workWhy not previously picked
NebulaGraphApache-2.0Distributed graph; mature; nGQLDistributed-server overkill for typical Crosswalker scale; nGQL is proprietary
JanusGraphApache-2.0Tinkerpop standard; pluggable backendJVM + Cassandra/HBase backend; operationally heavy
ArcadeDBApache-2.0Multi-model (document/graph/key-value/time-series/vector); MCP serverJVM-based; smaller community
MemgraphBSL 1.1Excellent Cypher; vector indexNot OSI open source
ArangoDBBSL 1.1 (since 2024)Multi-model; matureNot OSI open source
FalkorDBSSPLVector + Cypher + sparse-matrix engineSSPL — not OSI open source
GraphDB Free (Ontotext)Commercial (free tier)Best OWL reasoningCommercial license complications

6. Hybrid: layered Tier 3 (mirroring layered Tier 2)

Section titled “6. Hybrid: layered Tier 3 (mirroring layered Tier 2)”

What if Tier 3 mirrors Tier 2’s layered approach? E.g.:

  • Apache Jena Fuseki for RDF/SPARQL (canonical SSSOM endpoint)
  • DuckDB-on-server for SQL/tabular analytics (same engine as Tier 2)
  • Optional Apache AGE on Postgres for property-graph users with existing Postgres infrastructure
  • Optional TerminusDB as vault-mirror for git-style versioning

Required:

  • Operational complexity: running 2–4 server processes vs one
  • Federation across the layered components: Jena’s SERVICE clause + DuckDB-on-server’s HTTPFS + cross-engine joins
  • vs single-engine simplicity: at what user scale does layered Tier 3 become worth the operational cost?

Success criteria for the deliverable

Section titled “Success criteria for the deliverable”
  1. Engine evaluation matrix — Apache Jena Fuseki vs DuckDB-on-server+DuckPGQ vs TerminusDB-as-primary vs HelixDB vs layered Tier 3, scored on: license, governance, performance, RDF fit, SQL fit, graph traversal, ops complexity, multi-user concurrency
  2. Recommended Tier 3 default — single engine OR layered stack
  3. Migration path from AGE — for early adopters who started on AGE during the Ch 10 era
  4. Decision on AGE’s role — drop entirely, keep as fallback for Postgres-standardized environments, or re-affirm as default if alternatives have bigger problems

Out of scope

Section titled “Out of scope”
  • Re-evaluating Tier 2 (DuckDB-WASM + Oxigraph + Nemo) — that’s covered by Ch 14 (Grafeo evaluation) and the existing Ch 11 deliverables
  • TerminusDB’s vault-mirror role — already committed in TL;DR §2.6; this challenge could promote it to primary but the vault-mirror option stands either way
  • Implementation specifics — research only

Relationship to prior challenges

Section titled “Relationship to prior challenges”
  • Follow-on to Ch 11 — Ch 11 surfaced AGE concerns; this challenge acts on them
  • Coordinates with Ch 14 — Ch 14 is Tier 2-focused; this is Tier 3-focused
  • Independent of Ch 15 — different layer
Section titled “Related”