Choose Oracle Converged Features vs Specialized External Platforms Based on Workload Evidence

Learning outcomes

ServiceHub now has relational orders, JSON device payloads, XML partner envelopes, technician locations and relationship-heavy assignment data. The architecture team proposes five new specialist platforms—document, XML/content, GIS, graph and vector/search—before measuring whether Oracle's converged features already meet the workload. The opposite mistake is equally dangerous: forcing every workload into Oracle because “one database is simpler.” A good decision compares equivalent semantics, SLOs, scale and operational cost.

1. “Converged” means shared transactional/operational substrate—not one ideal engine for everything

Keeping related relational, JSON, spatial and graph views in Oracle can remove change-data-capture pipelines, duplicate identity/security policy, cross-system transactions, extra backup/DR runbooks and consistency lag. But a specialized platform can justify those costs when it materially improves a dominant workload or capability Oracle cannot meet economically.

2. Decision matrix

Chapter 25 will benchmark the vector-specific tradeoffs rather than treating “AI database” as a reason to skip evaluation.

3. Build a decision record in the database

BEGIN  EXECUTE IMMEDIATE 'DROP TABLE sh24_platform_eval PURGE';EXCEPTION WHEN OTHERS THEN  IF SQLCODE != -942 THEN RAISE; END IF;END;/CREATE TABLE sh24_platform_eval (  candidate VARCHAR2(40) NOT NULL,  criterion VARCHAR2(40) NOT NULL,  weight NUMBER NOT NULL CHECK (weight BETWEEN 1 AND 5),  measured_value VARCHAR2(200),  pass_flag CHAR(1)    CHECK (pass_flag IN ('Y','N')),  evidence_uri VARCHAR2(500),  PRIMARY KEY(candidate,criterion));INSERT INTO sh24_platform_eval(candidate,criterion,weight)VALUES('ORACLE_CONVERGED','TRANSACTION_CONSISTENCY',5);INSERT INTO sh24_platform_eval(candidate,criterion,weight)VALUES('ORACLE_CONVERGED','P95_LATENCY',5);INSERT INTO sh24_platform_eval(candidate,criterion,weight)VALUES('ORACLE_CONVERGED','P99_LATENCY',5);INSERT INTO sh24_platform_eval(candidate,criterion,weight)VALUES('ORACLE_CONVERGED','THROUGHPUT',4);INSERT INTO sh24_platform_eval(candidate,criterion,weight)VALUES('ORACLE_CONVERGED','FAILOVER_RTO_RPO',5);INSERT INTO sh24_platform_eval(candidate,criterion,weight)VALUES('ORACLE_CONVERGED','OPERATIONS_FTE_COST',3);INSERT INTO sh24_platform_eval(candidate,criterion,weight)VALUES('ORACLE_CONVERGED','LICENSE_INFRA_COST',3);COMMIT;

Add an identical criterion set for each external candidate. Leave measured_value empty until the prototype actually runs. A scorecard filled with opinions is not evidence.

4. Prototype contract: same semantics first

• Data: same representative cardinality, distribution, document/geometry/graph degree and update rate.
• Correctness: same filter/distance/path/consistency semantics and result checksum.
• Concurrency: same clients, request mix, think time and steady-state duration.
• Latency: median, p95, p99—not one stopwatch call.
• Throughput: successful requests/transactions per second.
• Resources: CPU, memory, I/O, network, index/storage footprint.
• Freshness: synchronous transaction or measured replication/index lag.
• Failure: node/process/network loss; retry/recovery correctness and RPO/RTO.
• Operations: deploy, patch, backup/restore, schema/index change, observability, on-call skill.
• Security: identity, authorization, encryption, auditing and secrets/key management across all copies.

5. Convergence has measurable integration value

If a JSON or graph query reads the same committed relational rows in the same database, there is no external CDC lag to measure. If a specialist platform receives a copy through Kafka/GoldenGate/custom pipelines, the architecture must measure end-to-end freshness and failure/replay behavior—not just the specialist engine's local query time.

6. External copies expand recovery and security scope

A second platform needs its own backup/restore, HA/DR, patching, certificates/secrets, RBAC, audit retention, capacity planning and incident response. It also introduces questions such as: after Oracle PITR, how is the external index/graph/document copy reconciled? Can consumers see data that Oracle rolled back? Who owns reindex/replay?

7. Use database-level measurements for the Oracle candidate

SELECT  SYSTIMESTAMP AS captured_at,  SYS_CONTEXT('USERENV','SERVICE_NAME') AS service_name,  SYS_CONTEXT('USERENV','CON_NAME') AS con_nameFROM dual;SELECT name,valueFROM v$sysstatWHERE name IN (  'user calls',  'execute count',  'user commits',  'session logical reads',  'physical reads',  'redo size')ORDER BY name;

Capture T1/T2 deltas around the prototype and pair them with client-side percentile latency. Database counters alone do not include external network/middle-tier time.

8. Instrument the candidate workload identity

BEGIN  DBMS_APPLICATION_INFO.SET_MODULE(    module_name => 'SH24_PLATFORM_POC',    action_name => 'ORACLE_CONVERGED'  );  DBMS_SESSION.SET_IDENTIFIER(    client_id => 'poc-run-001'  );END;/

Use equivalent labels in the external candidate's telemetry so traces/logs/metrics can be correlated by one run ID.

9. Deliberately wrong: choose a graph/document engine from a 1,000-row laptop demo

A tiny warm-cache test can hide network hops, replication lag, index build cost, failure recovery and operational load. It can also favor either system simply because one client driver warmed first. The repair is a representative dataset and concurrency ramp with the same acceptance criteria and deliberate failure/recovery tests.

10. Deliberately wrong: choose Oracle because “one system always costs less”

Convergence reduces integration surfaces, but a specialist platform can reduce engineering cost when a dominant capability is materially better—specialized GIS toolchains, graph algorithms, search relevance, document developer APIs or horizontal scale. The cost model must include both license/infrastructure and engineering/on-call complexity.

11. Go/no-go decision table

12. Licensing belongs inside the benchmark record

The current 26ai matrix lists Oracle Spatial and Graph and Property Graph/RDF technologies across all listed offerings, including Free. Other capabilities have distinct offering/option boundaries (for example parallel features, partitioned spatial indexes, management packs, Advanced Compression or RAC). An external product can also have per-node/core/ingest/storage licenses. Record the exact production entitlement—not merely what worked in a developer lab.

13. Rollback architecture before platform introduction

• Can the external index/cache be rebuilt completely from Oracle?
• Can dual writes be disabled without losing authoritative data?
• What is the source of truth during partial outage?
• How are schema/event-version incompatibilities handled?
• How long can consumers tolerate stale/unavailable derived results?
• What is the decommission path if the prototype fails its gates?

14. Cleanup

BEGIN  DBMS_APPLICATION_INFO.SET_ACTION(NULL);  DBMS_APPLICATION_INFO.SET_MODULE(NULL,NULL);  DBMS_SESSION.CLEAR_IDENTIFIER;END;/DROP TABLE sh24_platform_eval PURGE;

15. Production judgment and chapter close

Choose a converged feature when it preserves one transactional/security/recovery truth and meets the workload economically. Choose a specialized system when measured capability/SLO/scale/operations gains exceed the cost of copying, synchronizing, securing and recovering another platform. The burden of proof applies in both directions.

Current Chapter 24 baseline is Oracle AI Database 26ai RU 23.26.3, SQL Developer 26.2 and SQLcl 26.2.1.222.1617. Mandatory JSON/XML/spatial/property-graph labs fit Free's local 2-CPU/2-GB/12-GB envelope; that envelope is not an enterprise capacity result. Native JSON needs COMPATIBLE >= 20; 26ai XMLType TBX default needs COMPATIBLE >= 23.0; no lab raises COMPATIBLE. Spatial/Graph are currently included across listed offerings; parallel/partitioned/HA/pack features keep their separate licensing. Chapter 25 now applies the same evidence discipline specifically to VECTOR data, similarity search, hybrid retrieval and Select AI.

Authoritative references

• Oracle AI Database Documentation — JSON/XML/Graph/Spatial — current 26ai feature guides
• Licensing Information — offering/option/pack boundaries
• JSON-Relational Duality Developer's Guide — converged relational/document model
• Spatial Developer's Guide — geospatial semantics/indexing
• Property Graph Documentation — SQL graph and optional graph analytics ecosystem