What changed
If you have read something here before, this is what has moved since. There is no publication date on this page and no weekly or monthly promise: it is regenerated from the underlying data every time the site is built, so it is accurate whenever you read it.
Corrections and revisions
Substantive changes only: things that would alter what you concluded last time. Typos and wording are not listed.
- Database EngineeringLearning Paths
Refocused on engineering rather than operations. The pooling, replication and high-availability stages moved out - they belong to running a database, which the Production Infrastructure path and the middleware and architecture references cover. In their place: dedicated stages for data modelling and normalisation, index design, the query planner, and schema evolution under traffic.
- Agent memoryAI
Separated checkpointing, archives and long-term memory; added provenance and retention design plus the funes example. Corrected MemGPT benchmark attribution and the scope of the Generative Agents ablations. Human review of this revision is pending.
- Multi-agent orchestrationAgent Engineering
Added adaptive routing, cascade and critique patterns from the HydraFusion research preview, a handoff contract and shared-budget diagnostics. Clarified that the historical token multiple is not a universal sizing rule. Human review of the addition is pending.
- Inference Cost StructureAI Economics
Page created. All figures cited to vendor pricing pages and dated at time of observation.
- EU AI ActAI Governance
Page created from the Act text, with dates and penalties verified article by article.
- PgBouncerMiddleware
Corrected the claim that prepared statements are unusable in transaction mode. They have been enabled by default since 1.24.0.
- PgBouncerMiddleware
Added the CVE-2025-2291 context for the VALID UNTIL check in auth_query, and a release-behaviour table for 1.24.0 onward.
Releases and end-of-life dates that have passed
The last 120 days, from published support data. Scheduled dates are on the timeline.
- MongoDB 9.0 released
- Prometheus 3.14 reached end of life
- Prometheus 3.15 released
- MariaDB 13.0 released
- dbt Core 2.0 released
- TimescaleDB 2.30 released
- Kubernetes 1.37 released
- Prometheus 3.14 released
- Qdrant 1.19 released
- MongoDB 8.2 reached end of life
- Prometheus 3.5 reached end of life
- InfluxDB 3.11 released
- InfluxDB 3.9 reached end of life
- Redis 8.10 released
- TimescaleDB 2.29 released
- dbt Core 1.12 released
- SQL Server 13.0-sp3-acp reached end of life
- SQL Server 13.0-sp3 reached end of life
- Apache Spark 4.2 released
- Prometheus 3.12 reached end of life
- etcd 3.7 released
- MariaDB 10.6 reached end of life
- Apache Airflow 3.3 released
- Apache Airflow 3.2 reached end of life
- Prometheus 3.13 released
- Kubernetes 1.33 reached end of life
- InfluxDB 3.10 released
- InfluxDB 3.8 reached end of life
- dbt Core 1.10 reached end of life
Pages we know are ageing
Due for review against their sources, or already overdue. Listed because a reader deserves to know which pages we are least confident are current.
- Build Versus Buydue in 42 days
- Context engineeringdue in 42 days
- Context Window Economicsdue in 42 days
- Cost of Retrieval Versus Fine-Tuningdue in 42 days
- Forward-Deployed Engineeringdue in 42 days
- GPU Capacity and Utilisationdue in 42 days
- Inference Cost Structuredue in 42 days
- EU AI Actdue in 42 days
- Generative AI and the GDPRdue in 42 days
- ISO/IEC 42001due in 42 days
- NIST AI Risk Management Frameworkdue in 42 days
- Data and Model Poisoningdue in 42 days
- Excessive Agencydue in 42 days
- Hidden Context Exposuredue in 42 days
- Improper Output Handlingdue in 42 days
- Misinformationdue in 42 days
- Sensitive Information Disclosuredue in 42 days
- Supply Chaindue in 42 days
- Unbounded Consumptiondue in 42 days
- Vector and Embedding Weaknessesdue in 42 days
- Prompt engineeringdue in 42 days
- Payment Ledger Designdue in 42 days
- PostgreSQL High Availabilitydue in 42 days
- RAG at Production Scaledue in 42 days
- Read/Write Splittingdue in 42 days
- Transaction Monitoring and Sanctions Screeningdue in 42 days
- Choosing a relational databasedue in 42 days
- Choosing a vector storedue in 42 days
- Pooling and failover topologydue in 42 days
- Retrieval or fine-tuningdue in 42 days
- AI Engineeringdue in 42 days
- Claude and Anthropic toolingdue in 42 days
- Production Infrastructuredue in 42 days
- "Attention Is All You Need" - what it actually saiddue in 42 days
- "Retrieval-Augmented Generation" - what it actually saiddue in 42 days
- MariaDBdue in 42 days
- MongoDBdue in 42 days
- MySQLdue in 42 days
- PostgreSQLdue in 42 days
- SQL Serverdue in 42 days
- etcddue in 42 days
- HAProxy for PostgreSQLdue in 42 days
- keepaliveddue in 42 days
- Patronidue in 42 days
- PgBouncerdue in 42 days
- Pgpool-IIdue in 42 days
- Prometheusdue in 42 days
- AI Red Teamingdue in 43 days
- Agent memorydue in 43 days
- Agent planningdue in 43 days
- RAGdue in 43 days
- RAG for Operational Reportingdue in 43 days
- PostgreSQL 17 Installation and Configurationdue in 43 days
- Evaluating LLM systemsdue in 45 days
What to watch
Changes in direction already visible in primary sources. Full reasoning on signals.
- Model tool access is standardising on one protocol
Integrations built once become portable across hosts, which removes a real cost. It also moves the security question: an MCP server is a tool surface reachable by any connected host, so its permissions are the blast radius, and a protocol that changed statelessness and deprecated sampling in one version is not yet stable enough to build without a migration plan.
- Damage from AI systems is shifting from what models say to what they do
Controls that inspect model output are aimed at the previous problem. The controls that matter now bound what a compromised model is permitted to do: per-action authorisation, human approval for anything irreversible, and credentials scoped to the task rather than the user.
- Prompt-injection defences that publish strong results are being broken by adaptive attacks
A published defence benchmark is not evidence that a defence holds. Treat injection as an architecture problem rather than a filtering one, and assume the filter will eventually be bypassed when deciding what the model is allowed to reach.
- Vector search is becoming a database feature rather than a separate system
For workloads already on a relational database, a dedicated vector store now needs a specific justification - scale, or a feature the extension lacks - rather than being the default. The trade-off moves to recall measurement, because approximate indexes change which rows come back.
- Long-standing database projects are changing how they version releases
Version-pinning policies and upgrade-path assumptions written against sequential numbering will not survive the change. Anything that infers support status from a version number needs revisiting.
- Automatic database failover puts a consensus store in the availability path
Adopting automatic failover means adopting etcd's disk requirements as your database's availability requirements. Putting etcd on shared or network storage is a common cause of instability, and the failure presents as unexplained database demotions.
- The standard on-premises HA primitive does not work on major cloud networks
A keepalived design lifted from a datacentre will fail on these networks, often intermittently rather than obviously. Use the platform's load balancer for the address-failover role and keep keepalived for physical networks.
- AI regulation has moved from principles to dated obligations with penalties
Governance work can no longer be scheduled against internal readiness. The classification question - whether a use falls in an Annex III area - now determines a documentation and logging burden with a deadline attached.