feat: opaque partition-routing hooks for remote partition subgraphs - #829
Open
adsharma wants to merge 5 commits into
Open
feat: opaque partition-routing hooks for remote partition subgraphs#829adsharma wants to merge 5 commits into
adsharma wants to merge 5 commits into
Conversation
…graphs Partitioned node tables store rows in partition subgraphs (<parent>_p<i>). In a distributed deployment those partitions may live on remote hosts. Add a plain function-pointer interface (common/partition_routing_hook.h) that a distributed wrapper installs process-globally before the first database is opened: - locate(): placement decision per (parent table ID, partition index) - onPartitionCreate/onPartitionDrop: lifecycle notifications - bindScan(): read interception via a PartitionScanSpec - insertRow()/insertChunk()/lookupRow(): write/lookup interception The engine keeps ownership of the partition function, the catalog metadata and all schema information; wrappers only own where partitions live. PartitionHandle is opaque. Every hook is optional: with no hooks installed (the default) the engine behaves exactly as before. Includes the design doc covering the seam inventory, invariants, rejected alternatives, and landing order.
- Catalog: notify the wrapper when partition subgraph entries are created (always, so wrappers learn about new partitions and decide placement) and dropped (claimed partitions only). Renames are not reported: PartitionRef is ID-based and IDs survive renames. - StorageManager: skip local table creation for claimed partitions - no local storage, WAL or checkpoint state exists for them. Checkpoint, rollback, and metadata-snapshot serialization skip claimed partitions as well.
Reads: when expandPartitionedNodeTables finds a claimed partition, it asks the wrapper for a PartitionScanSpec (table function + bind-data factory) and attaches it to an internal clone of the partition's catalog entry - keeping the parent's schema, table ID and partition lineage. Partitions routed to the same wrapper scan collapse into one substitute entry, so a fully-claimed parent scans through the existing single-entry table-function path. Mixed local/remote scans of one parent cannot be planned and are rejected at bind time with a clear error; a defensive check in planNodeScan guards other paths into the same restriction. Supporting changes: - NodeTableCatalogEntry::CreateBindDataFunc now receives nodeUniqueName so foreign-backed entries can name output columns <nodeUniqueName>.<prop> / <nodeUniqueName>._ID, matching what planner schema lookups expect (same convention the duckdb/postgres extensions use in their getBoundScanInfo). - Add setScanFunction/setCreateBindDataFunc setters used to build substitutes. - Cardinality estimator skips scan-function-backed entries like foreign tables.
Point writes (INSERT/MERGE-create): map_insert resolves partition targets from catalog truth instead of pattern entries, so substituted scan entries don't affect write routing. resolveTargetTable returns null for remotely routed partitions; insertRemotely ships the evaluated row through insertRow() and uses the wrapper-assigned nodeID. MERGE output materialization goes through lookupRow(). The partition-function computation stays in the engine. Bulk writes (COPY FROM / INSERT ... SELECT): NodeBatchInsert keeps the engine's per-row computePartitionIndexes, then dispatches each row run either to its local target or to the wrapper's insertChunk(); remote targets skip local PK index building, error handlers and finalize. NodePartitionWriteInfo now carries the parent table ID so executors can build PartitionRefs without re-deriving lineage.
Add api-level tests that install a mock distributed wrapper and verify: - lifecycle notifications fire; claimed partitions keep catalog metadata but get no local storage; drop notifies per partition - point inserts route to the wrapper sink, which assigns node IDs; reads are served through the wrapper's consolidated scan (foreign-backed entry) - COPY FROM routes row runs through insertChunk - mixed local/remote scans of one parent are rejected at bind time while writes still route by the partition function
adsharma
force-pushed
the
feat/partition-routing-hooks
branch
from
August 22, 2026 03:06
3c7b60e to
aa221c7
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Keeps ladybugdb embedded and distribution-agnostic while allowing partition subgraphs (
<parent>_p<i>) to live on remote hosts. Introduces one opaque function-pointer interface —common::PartitionRoutingHooks— that a distributed wrapper installs before the first database is opened. Every place the engine touches a partition subgraph consults the hooks first; with no hooks installed (the default) behavior is unchanged bit-for-bit.Design doc:
partition-routing-hook.md(seam inventory, invariants, rejected alternatives).Interface
locate(ctx, ref, &handle)(parentTableID, partitionIndex)and returns an opaque handleonPartitionCreate/onPartitionDropbindScan(ctx, ref, handle, &spec)PartitionScanSpec(table function + bind-data factory)insertRow/insertChunklookupRowKey invariant: the engine owns the partition function and all catalog metadata; wrappers only own where a partition lives. Placement is computable locally without RPC;
PartitionHandleis opaque and never inspected by the engine.Changes by commit
CreateBindDataFuncnow receivesnodeUniqueNameso foreign-backed entries can name output columns<node>.<prop>/<node>._ID(the convention duckdb/postgres extensions already follow). Mixed local/remote scans of one parent are rejected at bind time.insertRow(wrapper assigns nodeIDs); COPY/batch inserts keep the engine'scomputePartitionIndexesand dispatch row runs to local targets orinsertChunk; MERGE lookups go throughlookupRow.Known limitations (documented, follow-ups)
Testing
PartitionRoutingTest.*api tests pass locally (4/4).ddl~partitioned.*e2e suite passes; copy/insert/scan/attach e2e filters show no new failures (copy~copy_to_csv.CopyToInvalidCasefails identically on a clean tree).