Elasticsearch 9.5 takes PromQL to GA and puts a no-inverted-index storage mode in preview

Elastic's DevRel newsletter for August 2026 says version 9.5 of Elasticsearch and the Elastic Stack promotes native PromQL, the Dashboards API, Cases as Data, and natural language authoring to general availability, while introducing Columnar Mode, a VectorDB index mode, a multimodal semantic field, and Agent Builder tracing in technical preview ^[1][2]. The item with a budget attached to it is Columnar Mode, which according to Elastic stores each field once in the column store with no inverted index by default, producing a smaller storage footprint and what the company calls a faster foundation for analytical queries and long retention ^[3].

That default is the whole decision. Built on Columnar Mode, Columnar Logs is the first profile designed for log data: it keeps a single inverted index on the message field so full-text search stays fast, and stores every other field in columnar format ^[4]. Elastic says storage use is significantly lower than standard logs without changing the search experience ^[5]. Both modes are opt-in and existing indices are not affected ^[6], so nothing changes in a running cluster until someone chooses to change it.

The newsletter offers no benchmark, percentage, or workload description behind "significantly lower" ^[7]. For a log-heavy team, that leaves the interesting question unanswered: what the profile does to queries that filter or aggregate on fields other than message, given that those fields no longer get an inverted index by default ^[3][4]. Term filters on high-cardinality fields are exactly the sort of thing that looks free in a demo index and expensive at retention scale. The honest read is that Columnar Logs is a candidate for a shadow index on a representative slice of your own log volume, not a setting to flip on a tier you care about, and it is technical preview in any case ^[4].

On the GA side, native Prometheus and PromQL support means teams can point existing Grafana dashboards and Prometheus tooling at Elastic and run PromQL alongside ES|QL with no changes to their existing setup, per Elastic ^[8]. A migration tool for bringing Grafana and Datadog dashboards across is also GA ^[9]. That is the more immediately usable half of the release: it changes where queries run without changing how data is stored. The Dashboards API is listed as GA with no further detail in the newsletter ^[1], which is worth noting for anyone planning automation against it.

The rest of 9.5 is preview work aimed at reducing configuration. VectorDB index mode sets up a vector index from a single index.mode setting, applying defaults for quantization, merge policy, and cache loading ^[10]; auto-calibration for DiskBBQ sets quantization depth, preconditioning, and oversampling by analysing the actual vectors in the index ^[11]. The new semantic field extends semantic search to images, audio, video, and PDFs, generating embeddings at ingest and allowing mixed content types in one field, powered by the jina-embeddings-v5-omni model ^[12]. Agent Builder now records every LLM call, tool invocation, and reasoning step as OpenTelemetry data in Elasticsearch, with cost-threshold alerts buildable against the trace index, and supports human approvals that can be routed to tools such as Slack ^[13][14].

Counting the newsletter's own lists, 9.5 lands four GA promotions and four preview additions ^[15]. Watch whether Elastic publishes storage and query benchmarks for Columnar Logs before it graduates, and whether the message-only inverted index holds up against your real filter patterns rather than a synthetic one.