mirror of
https://github.com/lancedb/lancedb.git
synced 2026-11-13 14:16:54 +00:00
Jack YeandClaude Sonnet 5
de95ed3f6d
A broad documentation pass across the areas LanceDB Enterprise and Embedded diverge on most: **Data Curation** -- GPU Function execution, Slurm scheduling (and what Slurm is, for readers new to it), blob late materialization as its own page, joins, aggregation/profiling, sampling, near-duplicate detection, chunking, distributed shuffle, and IVF-partition clustering. **Feature Engineering** -- computed columns and materialized views, restructured to lead with the Function-based path and show the SQL form side by side with Python throughout; a single Auto Refresh page instead of duplicating scheduling details on every refresh page; adaptive execution; checkpointing; GPU and Slurm placement for refresh workers; a Training Cache page for the PyTorch data loader's automatic cache benefit. **Security & Governance** -- an Audit Log reference (schema, access, grants), Secrets promoted to an ordinary catalog object with their own SQL DDL and a vending/rotation guide, and privilege-table fixes verified against the authorization source. **Deployment & Maintenance** -- Horizontal Scaling and OpenTelemetry moved into a real Deployment area; Automated Cleanup folded into Intelligent Auto-Maintenance (every other maintenance decision already lived there); RTE Compaction and Advanced Conflict Resolution documented in depth. **Landing page** -- the Enterprise feature comparison table reorganized (several redundant or stale rows removed, several new areas added), and the "why Enterprise" bullets replaced with ones tied to a concrete page (Performance, Scale, Fault tolerance, Multicloud deployment, Simplify operations) instead of vague promises. Plus a full pass addressing lancedb-gatekeeper's review: a real test-safety bug (a quickstart test could silently overwrite and delete a deployment's own `frames` table), a fabricated SQL statement with no backing implementation, an actual dedup-guarantee bug in `take_row_ids`, and dozens of documentation-accuracy fixes verified against source rather than taken on faith -- each either fixed, or left with an explicit note on why it needs a feature or deployment owner's sign-off instead of being guessed at. 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_018PXMCbRH6i5w47TsjeEUMT --------- Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
69 lines
3.2 KiB
Plaintext
69 lines
3.2 KiB
Plaintext
---
|
|
title: "Training Cache"
|
|
sidebarTitle: "Training Cache"
|
|
description: "Connect the PyTorch data loader to LanceDB Enterprise and training reads automatically benefit from the plan executor's data page cache."
|
|
icon: "gauge-high"
|
|
keywords: ["training", "dataloader", "cache", "data page cache", "enterprise", "epoch"]
|
|
availability:
|
|
embedded: unavailable
|
|
enterprise: available
|
|
---
|
|
|
|
The [PyTorch data loader](/training/data-loading) reads training data the
|
|
same way any other query does -- there is no training-specific read path.
|
|
Point `StreamingDataset` or a `DataLoader` at a table opened through an
|
|
Enterprise connection instead of a local one, and every read it makes
|
|
flows through the same [plan executors](/enterprise/data-page-cache) that
|
|
serve every other query, with the same data page cache in front of object
|
|
storage. No training-specific configuration turns this on; it's a
|
|
consequence of connecting to Enterprise at all.
|
|
|
|
```mermaid
|
|
flowchart LR
|
|
DL["PyTorch DataLoader<br/>/ StreamingDataset"] --> T["Enterprise-connected table"]
|
|
T --> PE["Plan executor"]
|
|
PE --> C{"In data page cache?"}
|
|
C -->|hit| Fast["Served from memory<br/>or local disk"]
|
|
C -->|miss| O[("Object storage")]
|
|
O --> Fill["Fills the cache,<br/>served"]
|
|
```
|
|
|
|
This covers the ordinary tabular columns a query reads -- the data page
|
|
cache's unit is a fixed-size page of table data. A [Blob v2](/management/table#managed-blobs)
|
|
column's payload is fetched through a separate range-request path keyed
|
|
to its own descriptor, not through this cache, so a training table mixing
|
|
plain feature columns with Blob v2 images or video gets this benefit only
|
|
for the former. A custom collate function or transform that does its own
|
|
file or network I/O -- decoding an image, say -- is also outside this
|
|
cache; only reads that go through the table query path benefit.
|
|
|
|
## Why this matters for training specifically {#why}
|
|
|
|
Training reads the same rows repeatedly: once per epoch, often for many
|
|
epochs. The first epoch's reads behave like any cold cache -- a miss per
|
|
page, each one filling the cache on the way. Every epoch after that reads
|
|
the same pages again, so a dataset that fits the cache's budget serves
|
|
almost entirely from memory or local disk after the first pass, instead of
|
|
re-reading object storage every epoch. A local-file or Embedded
|
|
connection has no equivalent remote cache to warm this way -- its
|
|
locality comes from whatever the local filesystem or OS page cache
|
|
already does.
|
|
|
|
[Cache Prewarm](/enterprise/cache-prewarm) extends this further: a plan
|
|
executor can warm its cache at startup, before training even begins,
|
|
instead of paying the first-epoch cold-cache cost during the run itself.
|
|
|
|
## What to read next {#next}
|
|
|
|
<CardGroup cols={2}>
|
|
<Card title="PyTorch Data Loader" icon="boxes-stacked" href="/training/data-loading">
|
|
Connecting a DataLoader to an Enterprise table.
|
|
</Card>
|
|
<Card title="Data Page Cache" icon="database" href="/enterprise/data-page-cache">
|
|
How the cache a training read benefits from actually works.
|
|
</Card>
|
|
<Card title="Cache Prewarm" icon="fire" href="/enterprise/cache-prewarm">
|
|
Warming the cache before the first epoch instead of during it.
|
|
</Card>
|
|
</CardGroup>
|