August closed with a run of PostgreSQL ecosystem news that matters well beyond the release calendar: a billion-scale vector search proof point on Aurora, a new managed Postgres service built specifically for AI agents, EXPLAIN support in Aurora DSQL, and fresh licensing math for the Postgres-vs-SQL Server decision. For Houston teams running Postgres across energy, healthcare, legal, and logistics workloads, here is what changed and what to do about it.

Billion-scale vector search on Aurora PostgreSQL

AWS published a case study on August 26 detailing how CORTO, an AI-powered legal technology platform serving more than 10,000 law firms, runs semantic search over 7.6 billion vectors in a 46 TB Aurora PostgreSQL cluster using pgvector. CORTO makes 2.5 billion documents searchable with sub-second query latency — and it chose Aurora PostgreSQL with pgvector over purpose-built vector databases.

The architecture is the part worth stealing. CORTO hash-partitions its embeddings table by firm_id and builds one partial HNSW index per firm:

CREATE TABLE embeddings (id bigint, firm_id bigint, vector vector(384))
  PARTITION BY HASH (firm_id);

CREATE INDEX CONCURRENTLY idx_embeddings_vector_firm_1
  ON embeddings_p0 USING hnsw ((vector::halfvec(384)) halfvec_ip_ops)
  WITH (m = 16, ef_construction = 256)
  WHERE firm_id = 1;

Because every query filters on firm_id, partition pruning routes it to a single partition and a single per-tenant partial index — so latency is bounded by one firm’s data, not the whole cluster. Storing vectors as halfvec at 384 dimensions cuts index size roughly in half with negligible recall loss, and Aurora Optimized Reads keeps hot vectors cached on local NVMe. For any Houston team building multi-tenant RAG or document search — contract review in the energy corridor, clinical literature at the Texas Medical Center — this is the reference pattern.

Tiger Data’s Ghost: a Postgres service built for AI agents

Timescale Inc., the New York company behind TimescaleDB, launched Tiger Data on June 9: a managed PostgreSQL database service designed for AI agents, with a new offering called Ghost now generally available. CEO Ajay Kulkarni’s argument is that agents experiment constantly — spinning up temporary environments, testing alternatives, and “going wild” in ways that make conventional database provisioning expensive and dangerous.

Ghost attacks that with fast forking: an agent can create an exact, independent copy of a dataset in seconds, backed by a copy-on-write storage layer (Fluid Storage) where instances share underlying data blocks and you pay only for data that changes. Pricing is based on active compute rather than per-database licenses — “one database per agent or one database per task,” as Kulkarni put it. If your Houston team is building coding-agent or workflow-agent pilots against Postgres, this is worth a look for dev/test economics alone.

EXPLAIN plans land in Aurora DSQL

AWS documented EXPLAIN support in Aurora DSQL on June 16. DSQL is AWS’s serverless, PostgreSQL-compatible distributed database for active-active single- and multi-Region workloads, and the new guidance shows how to read its query plans — which follow familiar PostgreSQL conventions but surface distributed-execution details:

EXPLAIN
SELECT account_id, transaction_date, description
FROM transaction
WHERE transaction_date > '2024-06-01';

Output like Full Scan (btree-table) on transaction reflects DSQL’s storage engine rather than a traditional heap, and the post walks through indexing patterns (including CREATE INDEX ASYNC) for joins and point lookups. Teams running distributed, multi-Region Postgres-compatible workloads — retail, logistics, or IoT telemetry around the Port of Houston — now have the same tuning loop they know from standard Postgres: explain, index, re-measure.

The $0 vs $15K licensing math

Tech Insider renewed the license-cost comparison on June 4 with Microsoft’s own pricing: SQL Server 2025 Enterprise lists at $15,123 per two-core pack with an eight-core minimum, which puts a modest 16-core production server around $60,000 in licenses alone — plus Software Assurance at roughly 25% of license value per year. PostgreSQL’s license cost is zero, and 49% of developers report using it in the Stack Overflow developer survey. SQL Server’s counter is depth in the Microsoft estate plus a built-in vector type in the 2025 release; Postgres answers with portability, extensions, and no per-core tax. For mid-market Houston energy-services and healthcare IT shops still carrying SQL Server licenses, that gap is a migration business case by itself.

What Houston teams should do this week

1. Steal the CORTO pattern if you’re building multi-tenant vector search: hash partition by tenant, one partial HNSW index per tenant, halfvec for storage. It scales to billions of vectors on plain Aurora.
2. Prototype agent dev/test with fast-forking Postgres. Whether it’s Ghost or a fork-on-write snapshot workflow, agents need cheap, disposable databases.
3. Run EXPLAIN on your DSQL hot queries if you’ve adopted Aurora DSQL — and check the AWS post for the distributed indexing patterns.
4. Model the SQL Server exit. At $15K per two-core pack, a licensing review can pay for the migration.

None of this requires a version upgrade or a forklift. It’s the Postgres ecosystem doing what it does best: proving scale on open infrastructure, and giving Houston teams cheaper, faster options for the AI and data workloads already on the roadmap.