MARCH 2026

The 'AI DBA' Dilemma, Postgres Tuning with Honeycomb + pganalyze MCP Server in Action

Hi there,

Welcome! I wanted to start this month's newsletter on a topic that's been top of mind for many. My perspective on AI tools has shifted considerably over the past year — from "this doesn't really solve complex tasks for me" to "this actually works pretty well for certain use cases." But the more capable these tools become, the sharper one dilemma gets: you can hand off the work, but an AI agent won't be responsible at 2am when the database goes down. The 'AI DBA' framing is becoming a real marketing buzzword, and I think it's missing the mark in ways that matter when things go wrong in production. In today's article I dig into why, and what I think we should be doing instead.

The session we recorded last month gets into the practical side of that. I walk through a real investigation using APM tracing data from Honeycomb alongside pganalyze data — using Claude Code and the pganalyze MCP Server (now in early access) to drive the analysis, without needing direct access to production and risk exposing anything the LLM shouldn't see. Worth watching if you're curious what this kind of workflow looks like day-to-day.

Interested what's going on with Postgres? This month is the final commitfest (1 month review cycles in Postgres development) for Postgres 19. With feature freeze coming up in early April, now is the time the final features get shaped. More on what got merged in next month's newsletter!

As always, you can find the complete list of pganalyze product updates at the end, including the availability of cluster-aware indexing recommendation for those on pganalyze Enterprise Server.

New Article

The Dilemma of the 'AI DBA'

AI tools are getting better at database work, but aiming to have an 'AI DBA' is setting the wrong goal. In this article, we explore why handing off the work to an AI agent doesn't mean handing off responsibility, and what a more grounded approach looks like for teams running Postgres in production.

Demo Recording

Postgres Performance Tuning with Honeycomb, Claude Code, and the pganalyze MCP Server

In this session, we walk through investigations using Claude Code alongside Honeycomb and pganalyze: tracing slow requests from APM spans all the way to EXPLAIN plans, without exposing sensitive data or connecting AI tools directly to your database. To request early access to the pganalyze MCP Server feel free to reply to this email - we're doing a phased rollout & we'd happy to put you on the list!

Community News

Upcoming events

We're heading to two Postgres community events this month. Ryan will be presenting at Nordic PGDay in Helsinki on March 24 and pgDay Paris on March 26, covering how to build a practical observability framework for Postgres that cuts through the noise and actually helps you find the queries worth investigating. Let us know if we'll see you there! We'll also be at PG Data 2026 in Chicago this June — more details on talks there soon.

Product Updates

What's New in pganalyze

Index Advisor

• Support schema names with uppercase letters

  • Previously, when the referenced tables or functions are in the schema with uppercase letters, we were unable to run the query analysis of the query, therefore the Index Advisor was failing to run with such queries.
• Fix handling of fully qualified table names not in search path

  Previously the query analysis process (part of Index Advisor) did not correctly handle such cases, leading to errors because the referenced tables could not be found.

• Fix handling of table names that need quoting (uppercase, special characters, etc)

MCP Server & AI Tools

• GraphQL API: Add getQueryExplainFromTrace resolver to get EXPLAIN plans from "db.postgresql.plan" trace spans
  • When the pganalyze OpenTelemetry tracing integration is utilized, the pganalyze-collector will insert special links into the respective traces to link back to the full EXPLAIN plan in pganalyze.
  • These links can be easily followed with a browser, but can be challenging for AI tools to resolve.
  • To assist AI tools / MCP servers, provide a dedicated getQueryExplainFromTrace GraphQL resolver that takes the necessary arguments from the link (datname, timestamp, fingerprint) and returns an EXPLAIN plan.

Other Improvements & Bug Fixes

• Improve performance with many connections
  • When the connection count at a given time exceeds thousands, the page was very slow or sometimes unable to load. Now it supports rendering multiple thousands of connections.
• Query Details Page: Fix visual indication of times when query was not run, but other queries had activity
  • This was incorrectly showing as missing data (dashed grey area) due to a recent change, when it should have been showing as zero calls instead. This also avoids the graph becoming unreadable for infrequently run queries.
• Single Sign-On: Support SAML certificate update
  • When a SAML Identity Provider certificate is rotated or approaching expiration, it can now be updated directly from the Integrations page without needing to reconfigure the entire SAML integration.
  • A warning banner is shown when the certificate expires within 14 days (or has already expired).
• Improve server and database dropdown navigation to better support long server names
  • The server dropdown now dynamically sizes its width based on the longest server name, expanding up to the right edge of the database select. Primary and replica badges are visually distinct and always fully visible alongside server names.
• Ensure JWT tokens generated by application use integers for "exp" field
• The "exp" field was previously incorrectly serialized as a string in the JWT token, which is a violation of the JWT spec, and may cause issues with proxies between the application and the end-user browser.

pganalyze Collector v0.69.0

• Improve query text collection to reduce missing query texts
  • Add a fingerprint cache that retains known query IDs across collection cycles, increasing collector memory use by up to about 9 MB per monitored server whilst reducing CPU overhead
  • For data from pg_stat_activity (Connections page in pganalyze), this reduces the occurrence of <truncated query>
  • For data from pg_stat_statements (Query Performance page in pganalyze), this reduces the likelihood of "query text unavailable" by re-using previous query texts already stored on the server side
• PlanetScale: Paginate log fetching to ensure freshness
  • Previously, only a single page of up to 1,000 log entries was fetched, which proved insufficient for high-traffic servers. The collector now paginates through all available entries
• AlloyDB: Skip collecting replication stats to avoid errors
• Helm chart: Allow setting pod labels and pulling images by digest
• Syslog handler: Add UDP listener
• Add packages for Debian 13 (Trixie)

pganalyze Enterprise Server Release v2026.02.0

• pganalyze Enterprise Server release v2026.02.0 is now available and 

  does not require a database migration when coming from v2026.01.0

  • If upgrading from older releases, review the migration notes on the v2026.01.0 release notes

• Postgres 14+ is now required for the internal statistics database used by Enterprise Server

• pganalyze Index Advisor now analyzes Postgres clusters as a whole rather than treating each server in isolation, combining workloads from primary and replica servers to make unified index recommendations across the entire cluster

• This release changes some private GraphQL API fields

• The full list of updates is available in the release notes including more information on the GraphQL API changes at the end

Earn a $25 Now a $50! Gift Card for a Review on G2

Help us get more community feedback and qualify for a free gift card while supplies last.