Timeplus Enterprise 3.4

Key Highlights

For a detailed tour of the new features in this release, see What's New in Timeplus Enterprise 3.4.

Key highlights of the Timeplus 3.4 release include:

1. Tabby, the Timeplus Data Agent: a conversational data agent built into every workspace that explores your data, triages pipeline health, proposes and (with approval) runs DDL, writes its own Python/JavaScript UDFs, remembers facts across conversations, learns reusable "skills", runs standing instructions on a schedule, and connects to external tools via MCP. Works with OpenAI, Anthropic, and any compatible endpoint (including Amazon Bedrock).
2. App Framework and App Marketplace: install, upgrade and manage ready-made real-time solutions — pipelines and dashboards bundled together — as self-contained .tpapp packages, browsable from a built-in catalog or built and published yourself.
3. Core engine: new experimental index types: vector-similarity and full-text (inverted) indexes land in timeplusd for the first time, enabling semantic search and fast token search directly in the engine.
4. Smarter tiered storage: merge-before-move and move-grace holding substantially reduce the number of small objects written to a cold/S3 storage tier under continuous streaming ingestion.
5. Bounded primary-key index memory: a new per-stream cache limit evicts least-recently-used primary-key indexes, closing a memory growth path left open by 3.3's lazy index loading.
6. Distributed query correctness fixes: SETTINGS target_nodes now works for ordinary historical queries, count() is no longer overcounted on co-located shards, and SYSTEM STOP MERGES/MOVES now covers every shard of a multi-shard stream.
7. A long list of replication, checkpoint and commit-path reliability fixes closed out under sustained production load — Raft deadlock and recovery fixes, checkpoint barrier loss, failed-commit recovery and data-loss fixes, mutable stream fixes, and more.

Supported OS

Deployment Type	OS
Linux bare metal	x64 or ARM chips: Ubuntu 20.04+, RHEL 8+, Fedora 35+, Amazon Linux 2023
Mac bare metal	Intel or Apple chips: macOS 14, macOS 15
Kubernetes	Kubernetes 1.25+, with Helm 3.12+

Releases

We recommend using stable releases for production deployment. Engineering builds are available for testing and evaluation purposes.

3.4.1

Released on 10-02-2026. Installation options:

• For Linux or Mac users: Downloads
• For Docker users (not recommended for production): docker run -p 8000:8000 docker.timeplus.com/timeplus/timeplus-enterprise:3.4.1
• For Kubernetes users: helm install timeplus/timeplus-enterprise --version 14.0.0. This chart version does not yet expose a dedicated enableAgent toggle for Tabby — Tabby ships enabled by default at the appserver level, matching bare-metal/Docker installs; a chart-level toggle is tracked in helm-charts#32 and not yet merged.

Component versions:

• timeplusd 3.4.1
• timeplus_appserver 3.4.1
• timeplus_connector 3.1.0
• timeplus cli 3.1.1
• timeplus byoc 1.1.0-rc.0

Changelog

This release consolidates all timeplusd changes from 3.3.1 through 3.4.1, plus the first appserver release of Tabby, the Timeplus Data Agent, and the App Framework.

Timeplus Appserver

See What's New in Timeplus Enterprise 3.4 for the full tour. Highlights:

• Tabby, the Timeplus Data Agent — chat, SQL permission modes with human-in-the-loop approval, skills, cross-conversation memory, background tasks, MCP external tool connections, downloadable artifacts, and support for OpenAI/Anthropic-compatible models including Amazon Bedrock
• App Framework — app marketplace/catalog, install/upgrade/uninstall lifecycle, per-app resource and dashboard management, and tooling to build and publish your own apps

timeplusd — Features and Enhancements

• Vector-similarity and full-text (inverted) index experiments (#12292)
• Merge expired parts before no-merge TTL moves (#12269)
• Hold fresh lone parts before no-merge TTL moves — ttl_move_grace_seconds (#12360)
• Bound lazy-loaded primary key index memory — primary_key_cache_max_bytes (#12284)
• Support historical table query on a specified node via SETTINGS target_nodes for ordinary queries, as well as for log external streams and system.timeplusd_log (#12344)
• Restrict trivial count() to the shards requested by the query (#12357)
• Make SYSTEM STOP MERGES/MOVES cover every shard of a multi-shard stream (#12304)
• Apply merge selector limits to OPTIMIZE STREAM PARTITION without FINAL (#12367)
• Reject read-only and write-once disks in CREATE STORAGE POLICY (#12377)
• Stream tool: partition-scoped backup and additive fill restore (#12331)
• Harden timeplusd meta CLI and add offline database drop (#12300)
• Add consume schema strategy (single/all/raw) to decode Confluent messages with schema (#12294)
• Upgrade pulsar-client-cpp to v4.2.0 (#12281)
• Upgrade contrib/avro to decode negative array block counts (#12364)
• Use Iceberg database storage_endpoint for S3 URL (#12356)

timeplusd — Bug Fixes

• Reject pausing a materialized view before its first pipeline build (#12328)
• Fix streaming arg_min/arg_max compile error for bfloat16 (#12334)
• Fix file writer finalize (Kafka PEM file write) (#12335)
• Fix checkpoint barrier loss in RemoteSource async-read path (#12326)
• Fix container-overflow in MergeTreeDataPartWriterOnDisk::cancel() after skip indices are finalized (#12369)
• Reset exec_mode for parallel-replica remote legs so materialized views over log external streams can start (#12373)
• Rebuild secondary index for primary-key-subset index key columns (#12348, #12351)
• Make NativeLog fetch honor the caller's byte budget (#12350)
• Recover trimmed Raft hard state from the durable checkpoint record (#12382)
• Fix Raft scheduler self-deadlock on outbound send backpressure (#12417)
• Add negate method to AggregateFunctionIf for streaming retract semantics (#12243)
• Keep storage policy tracked as in-use until deferred stream drop completes (#12390)
• Fill absent _tp_sn so coalesced mutable inserts commit _tp_time correctly (#12391)
• Treat remote() hosts that resolve to this server as local again (#12395)
• Rewind idempotent keys when an inline historical commit fails (#12392)
• Fix CREATE STREAM ... AS SELECT with an explicit column list, including a mutable-stream segfault (#12394)
• Retry a failed historical commit instead of wedging the committed sequence number (#12404)
• Refresh node memory total per heartbeat so runtime cgroup limit changes propagate (#12410)
• Fix null idempotent keys after a schema switch inside a historical commit batch (#12406)
• Fix NULL handling in nullable arg_min/arg_max: checkpoint recovery CORRUPTED_DATA and NULL values in the state (#12407)
• Release historical parts once a streaming query has read them (#12375)
• Parse creation query AST every time on provisioner retry (#12325)
• Hide stray Parquet symbol (#12421)
• Allow large operator-new allocations to throw MEMORY_LIMIT_EXCEEDED instead of exhausting server memory unchecked (ported from upstream) (#12355)

Not included: #12358 (New Analyzer/Planner, part 1) and #12383 (Iceberg S3 table support, part 1) — both are internal ports, not user-visible or enabled in 3.4.1. use_analyzer defaults to false; the new analyzer is opt-in only (SETTINGS use_analyzer=true) and not exercised by default. #12383 is a small HTTP-header fix inside the existing (pre-3.4) Iceberg database/external-stream feature, part of an in-progress S3 Tables write-support backport (#12379, still open) — it doesn't add a new capability on its own. Dropped from the customer-facing changelog; tracked internally instead.

Not included: #12321 (Decode pip output as UTF-8) — this is a forward-port of a fix (#12293) already released in 3.3.1; the 3.3 branch received the fix directly and it never made it back to develop until now. There is no new 3.4 behavior here — 3.3.1 and 3.4.1 behave identically. Omitted as a customer-facing "fixed in 3.4" line to avoid implying a regression that doesn't exist.

Upgrade notes

See the full Upgrade Notes table in What's New for details. In short:

• Tabby is enabled by default but inert until an admin configures an LLM endpoint; set enable-agent: false in the appserver's config.yaml to hide it entirely.
• If you configured Tabby before engineering build 3.4.11, check the "Max tokens" setting — it may still be at the old default of 1024.
• Set NEUTRON_ENCRYPTION_KEY before saving Tabby's LLM credentials in production; re-enter the key once if you set this variable after already saving it.
• Changelog materialized views using a nullable argument with sum_if/count_if/other _if aggregates need to be recreated after upgrading (checkpoint layout changed, #12243).
• If you ever ran MATERIALIZE INDEX ... WITH CLEAR on a mutable stream whose secondary index key is a subset of the primary key, re-run it once after upgrading to purge bad entries written by the old bug (#12351).
• Pulsar external streams must use a pulsar+ssl:// URL to request TLS; a plain pulsar:// URL with TLS settings no longer silently enables TLS.