Skip to content

perf(api): normalize import batches in bounded parallel windows #1497

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
DecisionNerd merged 7 commits into from
Sep 19, 2026
+1,517 −87
Merged

perf(api): normalize import batches in bounded parallel windows #1497

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
1677dca
perf(api): normalize original import batches in bounded private-pool …
DecisionNerd Sep 19, 2026
e5b02c2
test(api): qualify parallel normalization and small-host fixtures
DecisionNerd Sep 19, 2026
150f571
docs(api): record matching S18 fixed-run digests
DecisionNerd Sep 19, 2026
ed3a843
Merge remote-tracking branch 'origin/main' into perf/1472-parallel-no…
DecisionNerd Sep 19, 2026
5d900a3
fix(api): keep normalization work within the receipt contract
DecisionNerd Sep 19, 2026
43d00f4
docs: requalify normalization on corrected receipt contract
DecisionNerd Sep 19, 2026
bece15d
test(exec): bound ordered-map concurrency rendezvous
DecisionNerd Sep 19, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
196 changes: 109 additions & 87 deletions crates/graphforge-api/src/import_session.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,8 @@ use uuid::Uuid;

use crate::{BulkInputKind, CancellationToken, GraphConstructionBudgets, GraphForge, OperationId};

mod normalization;

const FORMAT_VERSION: u32 = 1;
const SESSION_DIR: &str = "import-sessions";
const MANIFEST: &str = "manifest.json";
Expand Down Expand Up @@ -780,85 +782,86 @@ impl GraphImportSession {
if source.kind.input_kind() != input_kind || source.staged {
continue;
}
let mut batch_index = 0_u64;
for_each_source_batch(&session_root, &source, batch_rows, |batch| {
if batch_index < source.batches_staged {
batch_index += 1;
return Ok(());
}
if cancellation.is_some_and(CancellationToken::is_cancelled) {
return Err(cancelled());
}
let operation = import_batch_operation(
self.manifest.operation_uuid,
source.sequence,
batch_index,
);
let batch = normalize_batch(graph, operation, input_kind, &batch)?;
if batch.num_rows() == 0 {
batch_index += 1;
self.manifest.sources[source_index].batches_staged = batch_index;
self.manifest.sources[source_index].inflight_batch = None;
self.persist_manifest()?;
return Ok(());
}
let recovering = source.inflight_batch == Some(batch_index);
if !recovering {
self.manifest.sources[source_index].inflight_batch = Some(batch_index);
self.persist_manifest()?;
}
let chunk_id = format!("import-{:020}-{:020}", source.sequence, batch_index);
let region = RegionScope::named("append");
let started = CallStart::now();
let staged = match (input_kind, cancellation) {
(BulkInputKind::Node, Some(token)) => {
construction.append_nodes_with_cancellation(&chunk_id, &batch, token)
normalization::for_each(
graph,
&session_root,
&source,
batch_rows,
self.manifest.operation_uuid,
cancellation,
|index, batch| {
let mut batch_index = index;
if cancellation.is_some_and(CancellationToken::is_cancelled) {
return Err(cancelled());
}
if batch.num_rows() == 0 {
batch_index += 1;
self.manifest.sources[source_index].batches_staged = batch_index;
self.manifest.sources[source_index].inflight_batch = None;
self.persist_manifest()?;
return Ok(());
}
let recovering = source.inflight_batch == Some(batch_index);
if !recovering {
self.manifest.sources[source_index].inflight_batch = Some(batch_index);
self.persist_manifest()?;
}
(BulkInputKind::Node, None) => construction.append_nodes(&chunk_id, &batch),
(BulkInputKind::Edge, Some(token)) => {
construction.append_edges_with_cancellation(&chunk_id, &batch, token)
let chunk_id =
format!("import-{:020}-{:020}", source.sequence, batch_index);
let region = RegionScope::named("append");
let started = CallStart::now();
let staged = match (input_kind, cancellation) {
(BulkInputKind::Node, Some(token)) => construction
.append_nodes_with_cancellation(&chunk_id, &batch, token),
(BulkInputKind::Node, None) => {
construction.append_nodes(&chunk_id, &batch)
}
(BulkInputKind::Edge, Some(token)) => construction
.append_edges_with_cancellation(&chunk_id, &batch, token),
(BulkInputKind::Edge, None) => {
construction.append_edges(&chunk_id, &batch)
}
};
self.operation_timings
.append
.record(started, staged.is_err());
if staged.is_ok() {
RegionScope::record_work("rows", batch.num_rows() as u64);
}
(BulkInputKind::Edge, None) => construction.append_edges(&chunk_id, &batch),
};
self.operation_timings
.append
.record(started, staged.is_err());
if staged.is_ok() {
RegionScope::record_work("rows", batch.num_rows() as u64);
}
drop(region);
if let Err(error) = staged {
drop(region);
if let Err(error) = staged {
self.manifest.sources[source_index].inflight_batch = None;
let remaining = self
.manifest
.limits
.max_rejected_rows
.saturating_sub(self.manifest.progress.rows_rejected);
self.manifest.progress.rows_rejected = self
.manifest
.progress
.rows_rejected
.saturating_add((batch.num_rows() as u64).min(remaining));
self.persist_manifest()?;
return Err(error);
}
batch_index += 1;
self.manifest.sources[source_index].batches_staged = batch_index;
self.manifest.sources[source_index].inflight_batch = None;
let remaining = self
self.manifest.progress.rows_accepted = self
.manifest
.limits
.max_rejected_rows
.saturating_sub(self.manifest.progress.rows_rejected);
self.manifest.progress.rows_rejected = self
.progress
.rows_accepted
.saturating_add(batch.num_rows() as u64);
self.manifest.progress.peak_batch_rows = self
.manifest
.progress
.rows_rejected
.saturating_add((batch.num_rows() as u64).min(remaining));
.peak_batch_rows
.max(batch.num_rows() as u64);
self.update_construction_progress(&construction.progress())?;
self.persist_manifest()?;
return Err(error);
}
batch_index += 1;
self.manifest.sources[source_index].batches_staged = batch_index;
self.manifest.sources[source_index].inflight_batch = None;
self.manifest.progress.rows_accepted = self
.manifest
.progress
.rows_accepted
.saturating_add(batch.num_rows() as u64);
self.manifest.progress.peak_batch_rows = self
.manifest
.progress
.peak_batch_rows
.max(batch.num_rows() as u64);
self.update_construction_progress(&construction.progress())?;
self.persist_manifest()?;
Ok(())
})?;
Ok(())
},
)?;
self.manifest.sources[source_index].staged = true;
self.manifest.progress.files_pending =
self.manifest.progress.files_pending.saturating_sub(1);
Expand Down Expand Up @@ -1102,7 +1105,7 @@ fn for_each_source_batch(
root: &Path,
source: &SourceRecord,
batch_rows: usize,
mut consume: impl FnMut(RecordBatch) -> Result<(), GfError>,
mut consume: impl FnMut(Option<RecordBatch>) -> Result<(), GfError>,
) -> Result<(), GfError> {
let path = root.join("sources").join(&source.name);
let tracker = graphforge_filesystem::FileCacheReleaseTracker::default();
Expand All @@ -1115,12 +1118,12 @@ fn for_each_source_batch(
tracker.clone(),
)
.map_err(storage)?;
for batch in
ArrowFileReader::try_new(BufReader::new(reader), None).map_err(storage)?
{
consume(batch.map_err(storage)?)?;
}
Ok(())
consume_source_batches(
ArrowFileReader::try_new(BufReader::new(reader), None)
.map_err(storage)?
.map(|batch| batch.map_err(storage)),
&mut consume,
)
})();
finish_source_cache_release(result, &tracker, "Arrow source")?;
}
Expand All @@ -1133,20 +1136,40 @@ fn for_each_source_batch(
.with_batch_size(batch_rows)
.build()
.map_err(storage)?;
for batch in reader {
consume(canonicalize_parquet_batch(
source.kind.input_kind(),
&batch.map_err(storage)?,
)?)?;
}
Ok(())
consume_source_batches(
reader.map(|batch| {
canonicalize_parquet_batch(
source.kind.input_kind(),
&batch.map_err(storage)?,
)
}),
&mut consume,
)
})();
finish_source_cache_release(result, &tracker, "Parquet source")?;
}
}
Ok(())
}

/// Drain admitted earlier batches before a later decode error, while still
/// inside the source reader's cache-cleanup scope. None marks the drain point.
fn consume_source_batches(
batches: impl Iterator<Item = Result<RecordBatch, GfError>>,
consume: &mut impl FnMut(Option<RecordBatch>) -> Result<(), GfError>,
) -> Result<(), GfError> {
for batch in batches {
match batch {
Ok(batch) => consume(Some(batch))?,
Err(error) => {
consume(None)?;
return Err(error);
}
}
}
consume(None)
}

#[derive(Clone)]
struct ImportChunkReader {
file: std::sync::Arc<File>,
Expand Down Expand Up @@ -1422,7 +1445,6 @@ fn normalize_batch(
kind: BulkInputKind,
batch: &RecordBatch,
) -> Result<RecordBatch, GfError> {
let _region = RegionScope::named("normalization");
match kind {
BulkInputKind::Node => graph.normalize_import_node_chunk(operation, batch),
BulkInputKind::Edge => graph.normalize_import_edge_chunk(operation, batch),
Expand Down
Loading