Highlights on INFINI Pizza

In-place columns for hot counters and flags

Mon, 01 Jan 0001 00:00:00 +0000

In-place columns for hot counters and flags #

Fields that change constantly — view counters, stock levels, feature flags, ranking scores — used to pay a full document rebuild on every partial update. Marking a numeric or boolean field with "inplace": true gives it a dedicated typed column instead:

PUT /my-collection
{
 "properties": {
 "title": { "type": "keyword" },
 "views": { "type": "integer", "inplace": true },
 "score": { "type": "double", "inplace": true },
 "active": { "type": "boolean", "inplace": true }
 }
}

Partial updates on such fields become O(1) column writes — no fetch, merge, and re-index of the document. Atomic-style operators work directly against the column:

Runtime node settings and dynamic log level

Mon, 01 Jan 0001 00:00:00 +0000

Runtime node settings and dynamic log level #

Node-local knobs no longer require a restart. GET/PUT /_node/_local/settings exposes transient overrides over the pizza.yml values — search lane budgets, the slowlog threshold, the default search timeout, trash retention — and they take effect immediately:

# drop the slowlog threshold to 200ms on one node, effective immediately
curl -X PUT http://127.0.0.1:28000/_node/_local/settings \
 -H 'Content-Type: application/json' \
 -d '{"node.search.slow_threshold_ms": 200}'

The flagship: log.level flips live — no restart, no dropped requests — which turns “turn on debug logging on the suspicious node” into a one-liner while preserving target-specific directives like openraft=warn.

Collection-wide manual compaction

Mon, 01 Jan 0001 00:00:00 +0000

Collection-wide manual compaction #

Reclaiming disk no longer means waiting for the background cadence or poking shards one by one. One request runs a merge pass on every started primary copy of the collection, cluster-wide:

curl -X POST http://127.0.0.1:28000/my-collection/_compact \
 -H 'Authorization: Bearer pz_bootstrap_key'

The response reports the queued batch job per hosting node — job id plus how many local primary passes it covers — so a node-level failure fails the row, not the request:

FIRE storage explorer in the console

Mon, 01 Jan 0001 00:00:00 +0000

FIRE storage explorer in the console #

The web console (/_ui/) gained a FIRE storage explorer: drill from a node’s shards down to a single segment’s .fire files and see exactly where the bytes go — a per-field cost treemap of the columnar FIRE storage, next to the inverted index and stored-field payloads. In the screenshot above, the collection’s Storage tab sizes every field by its on-disk cost, each block split into storage categories — so the question “which field is eating my disk” is answered by the biggest rectangle on screen.

semantic_text — one-field semantic search

Mon, 01 Jan 0001 00:00:00 +0000

semantic_text — one-field semantic search #

Semantic search used to mean wiring two fields: a text field for the source, a dense_vector field with vector_options pointing back at it. The ES-style semantic_text type (Elasticsearch 8.15+ alignment) collapses all of that into one declaration:

PUT /my-collection
{
 "schema": {
 "properties": {
 "body": {
 "type": "semantic_text",
 "service": "zai",
 "model": "GLM-4.6V",
 "dims": 3072
 }
 }
 }
}

From there everything derives. Writes carry plain text — the vector is embedded server-side at write time, in bulk and after partial updates alike. Search needs no field names and no vectors:

Keystore hot-reload

Mon, 01 Jan 0001 00:00:00 +0000

Keystore hot-reload #

Rotating a secret no longer restarts the node. ${keystore:...} references re-resolve into the running configuration on demand: console edits trigger the reload themselves, and out-of-band changes (a CLI edit, a replaced mounted Secret) are one call away.

curl -X POST http://127.0.0.1:28000/_node/_local/keystore/_reload \
 -H 'Authorization: Bearer pz_bootstrap_key'

The answer says exactly what happened — key names only, values stay write-only:

{
 "entries": 3,
 "read_only": false,
 "references": 2,
 "changed": ["node.embedding.endpoints", "catalog.join_token"],
 "unchanged": 0,
 "failed": []
}

A rotated embedding API key applies to the next inference call; a rotated catalog.join_token to the next join attempt. A reference whose entry disappeared keeps its running value and lands in failed until the entry returns — reload never aborts the node the way a missing entry aborts boot.

Model features & the /_chat facade

Mon, 01 Jan 0001 00:00:00 +0000

Model features & the /_chat facade #

Every model in the AI-services registry can now declare what it is for and what it accepts — features (embedding or generation) and inputs (text or image). The two axes are orthogonal: “multimodal” is not a category of its own but an inputs list beyond text.

node:
 embedding:
 endpoints:
 - name: zai
 url: https://api.z.ai/api/coding/paas/v4/embeddings
 chat_url: https://api.z.ai/api/coding/paas/v4/chat/completions
 api_key: sk-…
 models:
 - embedding-3  # embedding + text (the defaults)
 - {id: GLM-4.6V, dims: 3072} # embedding, dims declared
 - {id: glm-4-flash, features: [generation]}  # text generation
 - {id: glm-4v, features: [generation], inputs: [text, image]}  # vision LLM
 - {id: clip-v2, features: [embedding], inputs: [text, image]}  # multimodal embedder

Omitted fields keep the historical defaults, so every existing plain-string listing still means embedding + text — nothing to migrate.

Cluster-wide recycle bin

Mon, 01 Jan 0001 00:00:00 +0000

Cluster-wide recycle bin #

Accidental deletes used to be a race against the sweeper: deleting a namespace only dropped metadata, and the data directories drifted into the trash WITHOUT a recovery snapshot — permanently unrestorable. The recycle bin is now a first-class, cluster-wide surface with a namespace-centric view, a pre-restore integrity check, and one-request restore:

# One cluster-wide list — every node's trash merged by item, grouped by
# namespace, with the delete-time snapshot summary of each item.
GET /_cluster/trash

Deleting a namespace captures a restore snapshot for every collection it contains, so the WHOLE namespace comes back as a unit — original schemas, shard UUIDs, partition routing and placements included: