LDBC Full Suite (spec)

Last updated: 2026-08-05

Status: Specification in GraphForge docs. Generators, drivers, and audited runs execute in an external scale harness — not GraphForge core CI. Convenience catalog loaders (graphforge.datasets) remain a planned extension (Datasets overview); this page is the workload contract, not a shipped loader.

Shared evaluation policy (spec vs harness boundary, completeness vs first-fail, evidence schema, pinned identity): Scale Evaluation. Size labeling: GSI. Disk-limited framing: Scale Limits.

The Graph Data Council (GDC) (formerly Linked Data Benchmark Council / LDBC) maintains the LDBC benchmark suite. Workloads remain branded “LDBC benchmarks” after the 2025 rename. This document inventories the full official suite as published on ldbcouncil.org/benchmarks and defines how GraphForge treats it at spec level.

---

Suite inventory (authoritative portfolio)

Per GDC’s published benchmarks page (verified 2026-08-05):

Benchmark	Focus	Workloads (current)	Primary data
SNB — Social Network Benchmark	Property-graph DBMS (OLTP + OLAP-style)	Interactive (v1 / v2) and Business Intelligence (BI)	Synthetic social network (Datagen)
Graphalytics	Graph analysis platforms	Six core algorithms + reference outputs	Standard graph datasets (+ optional Graph500-class graphs)
FinBench — Financial Benchmark	Distributed transactional graph DBMS	Transaction workload (Analytics workload: future work)	Synthetic financial graph
SPB — Semantic Publishing Benchmark	RDF / SPARQL stores	Media-publishing ontology workload	RDF datasets

GraphForge’s product surface is property-graph + Cypher / analyst verbs. For harness planning:

• In scope for GraphForge-facing runs: SNB Interactive, SNB BI, Graphalytics, FinBench Transaction (as the SUT allows).
• Suite completeness still lists SPB, but SPB is RDF/SPARQL — treat as out of GraphForge product path unless a separate RDF binding exists. Document SPB so “full suite” means the GDC portfolio, not a silent omission.

Historical note: graph-analytics work that once lived near early SNB drafts was delegated to Graphalytics as the official LDBC analytics benchmark.

---

1. Social Network Benchmark (SNB)

Home: ldbcouncil.org/benchmarks/snb
Specification: LDBC SNB specification (PDF)
Generator (current): ldbc_snb_datagen_spark
Docs / query repos: ldbc_snb_docs, Interactive / BI impls under github.com/ldbc

Shared dataset

Interactive and BI share a social-network schema (Person, Message/Post/Comment, Forum, Organisation, Place, Tag, TagClass, and typed relationships such as knows, likes, hasCreator, …). Exact attribute sets and cardinality rules are owned by the SNB specification — do not invent alternate schemas for “LDBC-compatible” claims.

Scale factors

SF is defined by serialized CSV size in GiB (not GSI Size Tags). Authoritative SF lists live in the current SNB specification PDF — disclose the spec version used in evidence (pinned identity).

Run class	Typical SF set	Notes
Engineering / harness smoke	0.003, 0.1, 0.3	Laptop-feasible; non-audited; preferred developer default
Engineering / dedicated harness	1, 3, 10	Common published comparison class (esp. SF1); still not GDC-audited unless commissioned
Large harness-only	30, 100, 300, 1000, 3000, …	Spec-published production factors; dedicated runners only
GDC audited certification	Per current auditing rules / auditor	Member-commissioned; full disclosure report — distinct from engineering green

Concern	Rule
Reproducibility	Same SF + generator version + serializer → same data
Initial vs updates	Interactive splits ~90% bulk load vs update streams (see spec)
GSI labeling	Count loaded entities → full GSI; see SF → GSI crosswalk
Disclosure	Evidence must record SF, Datagen commit/release, driver commit/release, serializer, and whether the run is engineering vs audited

Workloads

Workload	Intent	Query / op classes (spec-level)
Interactive	Transactional neighborhood + updates	Complex reads, short reads, inserts/deletes; v1 and v2 driver workflows
Business Intelligence (BI)	Aggregation- and join-heavy analytics	Complex analytical queries over large graph fractions; microbatches of insert/delete

Generators and drivers (external)

• Datagen (Spark): produce CSV (or other serializers) for a chosen SF.
• Interactive / BI drivers: official LDBC driver repos execute the query mix, schedule updates, and collect latency/throughput.
• GraphForge may only expose thin ingest helpers later; the driver remains outside core.
• Identity: every engineering or audited claim must disclose Datagen and driver commit/release plus SNB spec version — see Pinned generator / driver identity and the evidence schema.

Pass criteria / validation (spec-level)

Official audited runs follow GDC auditing rules (member-commissioned auditors, full disclosure report, fees — see GDC “Auditing process for LDBC SNB”). For GraphForge harness / engineering runs (non-audited), require at minimum:

1. Correctness: driver validation / substitution parameters match expected results for the SF (or documented golden outputs from the LDBC tooling).
2. Workload completeness: the declared query set for Interactive and/or BI is executed (no silent subset presented as “full SNB”).
3. Disclosure: SF, generator commit/version, serializer, hardware, resource policy, and whether updates were enabled.
4. Honest partial runs: if only Interactive SF0.1 is green, say so — do not imply BI or audited certification.

Retrospective GDC reviews called out partial SNB compliance; GraphForge docs must not encourage “SNB-like” marketing without naming the exact workload + SF.

---

2. Graphalytics

Home: ldbcouncil.org/benchmarks/graphalytics
Driver: ldbc_graphalytics

Industrial-grade benchmark for graph analysis platforms (Giraph, GraphX, GraphBLAS, …). Consists of:

• Six core algorithms (BFS, PageRank, weakly connected components, local clustering coefficient, community detection via label propagation, SSSP — exact names/parameters per the Graphalytics specification PDF).
• Standard datasets with reference outputs for validation (SURF/CWI repository).
• Optional large synthetic graphs (including Graph500-class sizes) where the Graphalytics rules allow.

GraphForge-facing dataset shortlist (ordered)

Harnesses targeting GraphForge should prefer this ordered shortlist unless a full Graphalytics standard-benchmark job composition is explicitly claimed. Sizes are approximate from the Graphalytics specification; re-profile → GSI after load. Full catalog remains authoritative at the GDC/Graphalytics homes.

Order	Dataset id	Class	≈ n	≈ m	Typical GSI band	Role
1	wiki-Talk	Real R1 (2XS)	2.39M	5.02M	GU-06-MD-…	First engineering dataset
2	kgs	Real R2 (XS)	0.83M	17.9M	GU-05-SM-…	Gaming / denser
3	cit-Patents	Real R3 (XS)	3.77M	16.5M	GU-06-MD-…	Knowledge / citation
4	dota-league	Real R4 (S)	0.06M	50.9M	GU-04-XS-… (n) / high E	Dense gaming
5	Graph500-22	Synthetic G22 (S)	2.4M	64.2M	GU-06-MD-D00	Bridge to Official Graph500 SCALE 22
6	Datagen-7.9-fb (or nearest Datagen-S)	Synthetic D7.9 (S)	~1.4M	~85.7M	GU-06-MD-…	Datagen family smoke
7+	com-Friendster, Graph500-24+, larger Datagen	XL+	—	—	07+	Harness-only; not default laptop

For a “full Graphalytics at dataset X” claim: run all six core algorithms on X with reference-output validation. For engineering green on the shortlist: complete algorithms the SUT implements via the official Graphalytics driver path and label omissions (do not imply six-algorithm coverage if only a subset ran).

Pass criteria / validation (spec-level)

1. Algorithm outputs validate against reference outputs within the specification’s tolerances.
2. All six core algorithms run on each declared dataset for a “full Graphalytics” claim at that dataset.
3. Report platform, parallelism, dataset ids, driver commit/release, and Graphalytics spec version exactly.

Relation to GraphForge: analyst verbs (PageRank, components, paths, …) may implement kernels that overlap Graphalytics algorithms, but a Graphalytics result claim requires the official driver/validation path in the external harness — not a one-off Python script.

Relation to GSI / Graph500: Graphalytics dataset sizes and optional Graph500-class graphs should be labeled with GSI after load. The Official Graph500 size ladder and Derived density matrix are separate tracks — see Scale Evaluation. Graphalytics is the algorithm workload track.

---

3. FinBench (Financial Benchmark)

Home: ldbcouncil.org/benchmarks/finbench
Specification: FinBench specification (PDF)

Targets financial scenarios (anti-fraud, risk control): high-degree hubs, edge multiplicity, asymmetric directed graphs, time-window and recursive path patterns.

Workload	Status (per GDC / spec)
Transaction	Defined — OLTP-style complex reads + continuous insert/delete
Analytics	Future work (not yet a required GraphForge harness lane)

Scale factors (Transaction)

Per FinBench specification (v0.2.0-alpha class; disclose the exact spec version used). SF ≈ serialized CSV GiB; default temporal window three years from 2020; default split 97% initial bulk / 3% incremental. Published SF set:

SF	≈ CSV size (published datasets page)	Engineering default?	Notes
0.01	~6 MB	Yes — smoke	Smallest published factor
0.1	~66 MB	Yes — laptop/harness
0.3	~202 MB	Yes — harness
1	~679 MB	Dedicated harness	Spec validation-class scale
3	~2 GB	Dedicated harness
10	~6 GB	Large harness	Spec notes audited Transaction runs at SF10

Entity counts per SF (accounts, companies, transfers, …) are in FinBench Appendix B — re-count after load and emit GD-… GSI. Pre-built dataset tarballs: FinBench datasets.

Pass criteria / validation (spec-level)

1. Use FinBench datagen + Transaction driver at a declared SF from the table.
2. Meet driver validation for the Transaction query set.
3. Disclose SF, datagen/driver commit or release, FinBench spec version, hardware, and consistency/write mode.
4. Do not claim “full FinBench” if only a read subset ran, or if Analytics is implied before GDC publishes it.

Profile loaded graphs with GD-… GSI (directed).

---

4. Semantic Publishing Benchmark (SPB)

Home: listed under GDC Benchmarks
RDF/SPARQL workload based on a media-publishing ontology.

GraphForge posture	Detail
Suite inventory	Included — part of the official LDBC portfolio (name it; do not omit silently)
Product path	Out of scope for Cypher/property-graph GraphForge — RDF/SPARQL only
Harness / SUT	Inventory-only — omit SPB from GraphForge SUT execution; no pass/fail lane
Evidence	When listing suite coverage, record spb: "inventory_only" (or equivalent) rather than green/red

---

Planned convenience API (not the suite)

Future graphforge.datasets helpers may load small SNB slices for tutorials. That is not an LDBC-compliant run. Example shape only:

from graphforge import GraphForge
from graphforge.datasets import load_dataset  # planned — not shipped in v0.5.x

gf = GraphForge()
load_dataset(gf, "ldbc-snb-sf0003")  # illustrative name

Official compliance requires the external harness + LDBC drivers.

---

References

• Graph Data Council
• Benchmarks portfolio
• SNB · Graphalytics · FinBench
• LDBC GitHub org
• Graph Scale Index — size axis + SF crosswalk
• Scale Evaluation — harness contract, evidence schema, pinned identity, completeness vs first-fail
• Scale limits — product envelopes

Related

• Dataset overview
• Scale Evaluation — Graph500 tracks