perf: accelerate MCP graph build/incremental updates and reduce resident memory (RSS) by HouMinXi · Pull Request #1139 · tirth8205/code-review-graph

HouMinXi · 2026-10-28T18:37:00Z

Summary

This PR resolves two major performance pain points when using code-review-graph via MCP and daemon mode:

Slow MCP graph building & incremental updates
High resident memory (RSS) footprint during idle MCP server / daemon runtime and full builds

Key Optimizations

1. Graph Build & Incremental Update Speed (`incremental.py`, `tools/build.py`, `communities.py`, `flows.py`, `graph.py`, `parser.py`)

Chunked single-transaction batch persistence (store_file_batch): Replaces per-file transactions (store_file_nodes_edges) during full_build and incremental_update with 200-file chunked batch commits.
Eliminate N+1 dependent queries (find_dependents): Uses INDEXED BY idx_edges_target_kind with batched UNION + JOIN SQL queries and reuses store.get_file_hashes() instead of per-file get_nodes_by_file() lookups (~1,530 ms -> ~50 ms on graph.py).
Fast no-op short-circuit in _run_postprocess & incremental_update: Skips 7 cross-file resolvers and full flow/community recomputation when files_updated == 0 and no stale files were removed, and fixes changed_files=[] falsy fallback to full rebuild.
Concurrent Tree-sitter grammar probe pre-warming (_prewarm_parser_probes): Pre-warms grammar probes with per-grammar locks before spawning worker pools.
Lightweight edge scan & $O(E)$ community cohesion (communities.py, flows.py): Replaces store.get_all_edges() (which deserialized json.loads(extra) 80,886 times) with a lightweight tuple query and computes community cohesion in a single $O(E)$ pass instead of $O(C \times E)$. Batches community_summaries, flow_snapshots, risk_index, and flow_memberships inserts with executemany.
Shared bare-endpoint resolution context & fast path normalization: Shares _bare_endpoint_context_cache between resolve_bare_call_targets and resolve_bare_tested_by_sources, filters cross-file resolvers by target language extensions, and adds a pure-str fast path to normalize_file_path.

2. Resident Memory (RSS) Reduction (`main.py`, `tools/init.py`, `graph.py`, `embeddings.py`, `parser.py`, `constants.py`, `daemon.py`)

PEP 562 lazy loading in main.py & tools/__init__.py: Defers loading heavy dependencies (networkx, igraph, tree_sitter_language_pack, parser.py) until a tool actually needs them, while preserving TYPE_CHECKING annotations and backward-compatible module attributes.
SQLite cache & temp table bounds (graph.py, embeddings.py): Caps SQLite page cache (PRAGMA cache_size=-2000, ~2 MB), guarantees DROP TABLE IF EXISTS for temporary impact-radius/unresolved-call tables in finally blocks, and adds GraphStore.shrink_memory() + release_os_memory() (gc.collect() + Linux malloc_trim(0)).
Parser & AST cache lifecycle (parser.py, incremental.py): Adds CodeParser.clear_caches() and flushes pending_batch every 200 files so full-build peak RSS drops by over 35%.
Streaming Top-K vector search (embeddings.py): Uses _iter_scores() generator with heapq.nlargest instead of materializing full-table score lists, and adds clear_model_cache().

Benchmark Results (378 files, 8,869 nodes, 81,885 edges)

Build & Incremental Speed

Scenario	Baseline (`d72e8fb`)	Optimized	Improvement
CLI Full Build + Full Postprocess (`ProcessPoolExecutor`)	`12,449 ms`	`5,689 ms`	2.19x faster (-54.3%)
MCP Non-TTY Full Build + Full Postprocess (`ThreadPoolExecutor`)	`29,956 ms`	`15,393 ms`	1.95x faster (-48.6%)
Incremental Update (no file change / short-circuit)	`3,848 – 4,288 ms`	`106 – 117 ms`	32.9x – 40.5x faster (-97%)
Incremental Update (single leaf file `main.py`, 35 deps)	`3,174 ms`	`813 ms`	3.91x faster (-74.4%)
Incremental Update (single core file `graph.py`, 186 deps)	`4,517 ms`	`1,089 ms`	4.15x faster (-75.9%)

Resident & Peak Memory (`VmRSS` / `VmHWM`)

Scenario	Baseline	Optimized	Reduction
`main.py` MCP Server Idle Startup (`VmRSS`)	`121.37 MB`	`72.42 MB`	-48.95 MB (-40.3%)
`code_review_graph.tools` Cold Import (`Delta RSS`)	`+29.49 MB`	`+1.19 MB`	-28.30 MB (-96.0%)
`GraphStore` Open Instance (`VmRSS`)	`38.49 MB`	`16.96 MB`	-21.53 MB (-55.9%)
Full Build (`postprocess='full'`) Peak Memory (`VmHWM`)	`164.83 MB`	`106.63 MB`	-58.20 MB (-35.3%)
Full Build + Postprocess + Bridge Nodes -> Close Store (`VmRSS`)	`144.38 MB`	`83.13 MB`	-61.25 MB (-42.4%)

Verification & Quality Gates

Graph Equivalence: 100% identical outputs (8,869 nodes, 81,885 edges, 139 flows, 19 communities, 8,803 FTS5 rows).
Ruff: uvx ruff check code_review_graph/ -> All checks passed! (0 errors)
Mypy: uvx --with types-networkx mypy code_review_graph/ --ignore-missing-imports --no-strict-optional -> Success: no issues found in 77 source files (0 errors)
Pytest: Full test suite passed (5045 passed, 0 failed).

github-actions · 2026-10-28T18:38:07Z

code-review-graph review

Overall risk: 0.85 (CRITICAL) — 136 changed function(s)/class(es), 86 affected flow(s), 5 test gap(s) (2 with no tested caller found, 3 reached only through a caller)

Risk-scored changes

Risk	Level	Symbol	Location	Tested
0.85	critical	code_review_graph/custom_languages.py::_validate_entry	code_review_graph/custom_languages.py:201	yes
0.85	critical	code_review_graph/graph.py::GraphStore._invalidate_cache	code_review_graph/graph.py:873	yes
0.75	high	code_review_graph/graph.py::GraphStore.get_impact_radius_sql	code_review_graph/graph.py:3063	yes
0.70	high	code_review_graph/constants.py::normalize_file_path	code_review_graph/constants.py:345	yes
0.65	medium	code_review_graph/constants.py::release_os_memory	code_review_graph/constants.py:368	yes
0.65	medium	code_review_graph/graph.py::GraphStore.refresh_target_resolution	code_review_graph/graph.py:2046	yes
0.65	medium	code_review_graph/graph.py::GraphStore._get_bare_endpoint_context	code_review_graph/graph.py:2345	yes
0.65	medium	code_review_graph/graph.py::GraphStore._resolve_bare_endpoints	code_review_graph/graph.py:2463	yes
0.65	medium	code_review_graph/graph.py::GraphStore._build_networkx_graph	code_review_graph/graph.py:3953	yes
0.65	medium	code_review_graph/incremental.py::_get_code_parser_cls	code_review_graph/incremental.py:40	yes

Affected execution flows

main — criticality 0.72, 1103 node(s) across 45 file(s)
benchmark_debug_workflow — criticality 0.70, 187 node(s) across 16 file(s)
run_token_benchmark — criticality 0.69, 57 node(s) across 6 file(s)
start_then_route — criticality 0.68, 24 node(s) across 5 file(s)
process — criticality 0.68, 142 node(s) across 16 file(s)
...and 81 more affected flow(s)

Test gaps

code_review_graph/graph.py::GraphStore.rollback (code_review_graph/graph.py:1282)
code_review_graph/main.py::run_fn (code_review_graph/main.py:261)

Reached only through a caller

No test names these directly, but a tested caller reaches them along the call path shown. That is a path in the call graph, not a record of execution -- the caller's tests may never take this branch. Treat it as where to look, not as coverage: these count as gaps and are scored as untested.

code_review_graph/embeddings.py::EmbeddingStore._search_pure_python (code_review_graph/embeddings.py:1530) — via code_review_graph/embeddings.py::EmbeddingStore.search, 1 hop(s)
code_review_graph/graph.py::GraphStore._get_test_helper_routes (code_review_graph/graph.py:1626) — via code_review_graph/graph.py::GraphStore.get_caller_test_routes, 1 hop(s)
code_review_graph/hcl_resolver.py::resolve_hcl_module_references (code_review_graph/hcl_resolver.py:15) — via code_review_graph/incremental.py::full_build, 2 hop(s)

Token savings: this graph-backed report used ~317,231 fewer tokens (~56%) than reading every changed file in full (estimated, chars/4 approximation).

Powered by code-review-graph — local-first analysis; no code leaves the CI runner.

Copilot

🟡 Changes recommended

Unresolved rebuild handling can leave derived graph data stale, while an igraph import failure can replace existing communities with an empty set instead of using the fallback.

6 open issues

What changed in this PR

This PR targets faster graph builds and incremental updates, plus lower resident memory use in MCP servers and daemons.

Changes:

Batches persistence and resolver queries, skips unchanged-graph work, and prewarms parser probes.
Defers heavy imports, adds cache cleanup, and streams top-K embedding search.
Adds direct helper tests and cleans up existing test formatting.

File	Description
tests/test_transactions.py	Cleans up imports and formatting.
tests/test_prompts.py	Wraps a long assertion string.
tests/test_parser.py	Exercises cache cleanup and unreachable-call helpers.
tests/test_main.py	Exercises lazy attributes, offloading, and prompts.
tests/test_incremental.py	Exercises build, resolver, and daemon helpers.
tests/test_graph.py	Exercises batching, resolution, and memory cleanup.
tests/test_fts_sync.py	Cleans up imports and formatting.
tests/test_embeddings.py	Tests model-cache cleanup and score streaming.
tests/test_custom_languages.py	Exercises validation and lazy grammar imports.
tests/test_communities.py	Exercises cohesion and community-detection helpers.
code_review_graph/tools/build.py	Adds no-op handling and batches summary writes.
code_review_graph/tools/__init__.py	Lazily loads exported tool implementations.
code_review_graph/temporal_resolver.py	Limits call queries to Java files.
code_review_graph/spring_resolver.py	Limits call queries to Java files.
code_review_graph/scoped_resolver.py	Batches queries for relevant language files.
code_review_graph/rescript_resolver.py	Adds an early exit and batches updates.
code_review_graph/python_resolver.py	Reduces JSON parsing and adjusts update transactions.
code_review_graph/parser.py	Optimizes traversal, probe locking, and cache cleanup.
code_review_graph/main.py	Defers tool, prompt, and graph imports.
code_review_graph/incremental.py	Batches writes and optimizes dependency lookup and parsing.
code_review_graph/graph.py	Adds memory controls and caches resolution context.
code_review_graph/flows.py	Batches flow memberships and short-circuits entry checks.
code_review_graph/event_resolver.py	Narrows derived-event cleanup queries.
code_review_graph/embeddings.py	Streams top-K scores and adds memory cleanup.
code_review_graph/daemon.py	Reclaims memory after initial builds.
code_review_graph/custom_languages.py	Defers grammar-package imports.
code_review_graph/constants.py	Adds shared path and memory-reclamation helpers.
code_review_graph/communities.py	Defers igraph loading and uses lightweight edge records.
code_review_graph/analysis.py	Releases graph caches after bridge-node analysis.

_{Review effort: Balanced}

HouMinXi · 2026-11-05T17:48:37Z

+    global ig
+    if ig is None and IGRAPH_AVAILABLE:
+        try:
+            import igraph as _ig  # type: ignore[import-untyped]
+
+            ig = _ig
+        except ImportError:
+            pass
    if ig is None:
        return []


Fixed in 28f7bb7: _load_igraph() clears IGRAPH_AVAILABLE when the import raises, and every caller now falls back to file-based grouping for that call.

HouMinXi · 2026-11-05T17:48:41Z

+    use_incremental = (
+        not full_rebuild
+        and changed_files is not None
+        and (bool(changed_files) or not build_result.get("stale_files_removed"))
+    )


Fixed in 28f7bb7: use_incremental now requires a non-empty changed_files; an empty hint takes the full recompute path, and genuine no-ops stay on the existing branch.

HouMinXi · 2026-11-05T17:48:58Z

            conn.execute("PRAGMA journal_mode=WAL")
            conn.execute("PRAGMA busy_timeout=5000")
+            conn.execute("PRAGMA cache_size=-2000")
+            conn.execute("PRAGMA temp_store=MEMORY")


Fixed in 28f7bb7: PRAGMA temp_store=MEMORY is gone, so impact-radius working tables stay file-backed and shrink_memory can reclaim them.

HouMinXi · 2026-11-05T17:49:02Z

+        cached = getattr(self, "_bare_endpoint_context_cache", None)
+        if cached is not None:
+            return cached


Fixed in 28f7bb7: upsert_node, upsert_edge and commit now clear _bare_endpoint_context_cache, so resolution never reuses evidence a new edge invalidated.

HouMinXi · 2026-11-05T17:48:44Z

+        else:
+            grammars.add(lang)


Fixed in 28f7bb7: rescript, vbnet, yaml and properties are excluded from the prewarm set; a test asserts only the grammar-backed language is probed.

HouMinXi · 2026-11-05T17:48:48Z

+    impl_candidate = tool_name[:-5] if tool_name.endswith("_tool") else tool_name
+    for _cand in (impl_candidate, f"{impl_candidate}_func", tool_name):
+        if _cand in _LAZY_TOOL_NAMES and _cand not in globals():
+            _resolve_attr(_cand)
+            break
+    if provenance and "with_provenance" not in globals():
+        _resolve_attr("with_provenance")
+
+    run_fn = (
+        (lambda: _resolve_attr("with_provenance")(work(), root))
+        if provenance
+        else work
+    )


Fixed in 28f7bb7: lazy resolution moved into the worker callable, so the cold import runs off-loop and inside the timeout.

HouMinXi · 2026-11-05T18:25:50Z

Addressed the review in 99fd19c. Review of that change surfaced three further defects, now fixed in the same commit:

The incremental helpers join on nodes.file_path, which is absolute, while update inputs are repository-relative. Every incremental flow and community pass therefore matched nothing. The hint now reaches both helpers in the spelling the graph stores.
A broken native install raises OSError, not ImportError, and an importable-but-empty package raises nothing at all. Both escaped the fallback and failed the postprocess stage; the loader now handles both, logs the cause, and checks that the type it needs exists.
PRAGMA temp_store is pinned to FILE so the guarantee does not depend on the SQLite build default, and rollback() drops the resolution context like commit() does.

Tests: the fallback, absolute-path handoff, cache drops and prewarm set each have a test verified to fail when its fix is reverted. Suites touched: 366 passed, 4 skipped.

HouMinXi · 2026-11-06T12:32:32Z

Extended in 23c1021 with the structural leftovers: the resolvers now commit through the store instead of the raw connection, so cache invalidation is no longer bypassed; a context fill that races an invalidation publishes only if the cache epoch is unchanged; custom languages are probed under their grammar name; and the networkx graph is built outside the cache lock and published under it. 316 passed, 4 skipped across graph, flows, incremental, communities and the resolvers.

…ent memory (RSS)

…ized symbols

HouMinXi · 2026-11-07T16:12:39Z

Rebased onto staging (d72e8fb). The branch is three commits again (9c0f5de, 20eae3f, 3fca15a) and the tree is clean against ruff.

The rebase dropped several parts of this PR because staging now carries an equivalent or better version:

Embedding search: staging scores with numpy and documents the vector cache (perf(embeddings): score stored vectors with numpy, not one component at a time #1090). This branch's embedding change is gone rather than duplicated.
Watch lifecycle: staging starts the watcher with run_post_processing and joins it before closing the store (fix(mcp): run post-processing during auto-watch updates #1108, fix(mcp): join auto-watch before closing its GraphStore #1127). Kept staging's loop; the branch's lazy-import variant is gone.
Incremental path bridging: incremental_trace_flows now expands relative paths with repo_root (Incremental flow re-detection never finds new entry points (relative vs absolute path mismatch) #569, fix(flows): rematch incremental flows across node replacement #673), so the duplicate hint this branch passed (_fts_file_hint) is removed, along with its test.
Write batching in the parse loop: staging routes writes through store_parsed, which captures flow entry points before writing. Kept that path; the batch writer and the pre-fetched hash lookup are gone.

What the branch still adds on top of staging:

igraph fallback that also survives OSError and an importable-but-empty package, with the availability flag honoured.
temp_store=FILE pinned rather than left to the SQLite build default.
Cache invalidation on upsert_node, upsert_edge and commit; resolvers commit through the store; a fill that races an invalidation does not publish.
Prewarm skips languages with no bundled grammar and probes custom languages under their grammar name.
identity_rebuild never takes the incremental branch, and lazy tool resolution runs inside the worker.

Tests: 389 passed, 4 skipped across graph, flows, incremental, communities, main and the fallback tests; the weighted impact scoring suite passes 19/19.

The build/RSS pass could leave the graph stale or block the event loop on six paths: - A lazy igraph import that failed after find_spec() succeeded kept the availability flag set, so community detection returned an empty set and replaced existing communities instead of using file-based grouping. - An empty change hint from the C++ identity rebuild took the incremental path, whose flow and community helpers return immediately for an empty list, so derived data stayed stale. - PRAGMA temp_store=MEMORY kept the impact-radius working tables resident after the delete, past the point where shrink_memory can reclaim them. - The bare-endpoint resolution cache was not cleared by upsert_node, upsert_edge or commit, so a later resolution reused evidence that a new IMPORTS_FROM edge had already invalidated. - Grammar prewarming probed rescript, vbnet, yaml and properties even though those languages are parsed by fallbacks and never load a grammar. - Lazy tool resolution ran inside the async wrapper before the worker was dispatched, blocking the loop on a cold import outside the timeout. Review of the first pass surfaced more: - The incremental helpers join on nodes.file_path, which is absolute, while update inputs are repository-relative, so every incremental flow and community pass matched nothing. The hint now reaches both helpers in the spelling the graph stores. - A broken native install raises OSError rather than ImportError, and an importable but empty package raises nothing at all; both escaped the fallback and failed the postprocess stage. The loader now handles both, logs the cause, and verifies the type it needs is present. - temp_store is pinned to FILE so the guarantee does not rest on the SQLite build default. - The resolvers committed through the raw connection, bypassing cache invalidation; they now commit through the store. rollback() drops the resolution context like commit() does. - A context fill that raced an invalidation could publish stale evidence; it now publishes only if the cache epoch is unchanged. - Custom languages are probed under their grammar name, not their alias. - The networkx graph is built outside the cache lock and published under it, so a long read no longer stalls writers. Tests cover the fallback, the absolute-path handoff, the cache drops, the prewarm set and the custom-language mapping; each was verified by reverting its fix and watching the test fail. Signed-off-by: Minxi Hou <houminxi@gmail.com>

HouMinXi · 2026-11-07T16:59:10Z

The 3.10/3.11/3.12 failures after the rebase were mine, and they were real: _load_igraph checked hasattr(module, "Graph") outside the handler, so a native build that imports and then raises OSError on first attribute access escaped on those interpreters. The type is now touched inside the same try with AttributeError added to the handler. Reproduced on 3.12 locally (OSError: libigraph.so: cannot open shared object file), fixed, and the same six suites pass there: 389 passed, 4 skipped. ae44bd2.

…1139) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…ew of tirth8205#1139) - The bare-call evidence context was cached until some write cleared it, but resolvers write through store._conn and skip that. A new import added that way was never seen: the next pass left the call bare. The call pass now always builds it fresh, and the tested-by pass drops it, so it never outlives the pair or holds every node name in a watch store. - GraphStore.close() and EmbeddingStore.close() ran gc.collect(). Every MCP tool call closes a store, and a full collection costs 10-15 ms in the server process, so it moved to the end of a build only. - EmbeddingStore set temp_store=MEMORY, the opposite of the graph fix. - Removed the batch writer leftovers: pending_batch was never filled and _can_batch_store was never called. - _single_hop_dependents and detect_communities sniffed whether a test had patched a GraphStore method. The dependents query is now GraphStore.get_direct_dependent_files, and the community edge scan always runs. - parser.py had its own copy of normalize_file_path; it re-exports the one in constants.py. - Restored the comments the PR deleted in the evidence builder. Each behaviour has a test in tests/test_derived_data_fallbacks.py that fails on the PR head. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

tirth8205#1139) main() opens the watch store with GraphStore(...), but the PR imported GraphStore only under TYPE_CHECKING and served it through the module __getattr__, which a bare global lookup never reaches. The incremental module (and so graph.py) loads with the server anyway, so GraphStore and find_project_root are imported normally again; only tool and prompt modules stay deferred. The worker also guessed which implementation to load from the tool name. It now loads exactly the names the tool's lambda reads, and a test fails if a deferred name is read anywhere the worker does not resolve it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

tirth8205#1139 skips flows and communities when an update stored no file. tirth8205#964 rebinds C# calls on a project-only change and marks the flows dirty, so the skip left them stale. A dirty mark now ends the skip. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

tirth8205#1139 loads the tool modules on first use. In test_cli.py an update test patches GraphStore and get_db_path while they load for the first time, so the mocks stayed bound in them and the real search test from tirth8205#823 that ran next found nothing. The test conftest now loads them up front. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

tirth8205#1139 makes graph.py load the parser on first use, so EdgeInfo is only a type name there. tirth8205#715 builds EdgeInfo objects when it aligns edges with colliding names, which raised NameError on every store. The two methods now import it themselves. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

tirth8205#1139 loads igraph on first use through _load_igraph, so setting IGRAPH_AVAILABLE alone no longer makes detect_communities take the Leiden path when igraph is not installed. The tirth8205#882 test now stubs the loader too. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

tirth8205 requested a balanced review from Copilot November 5, 2026 17:26

Copilot started reviewing on behalf of tirth8205 November 5, 2026 17:27

View session

Copilot AI reviewed Nov 5, 2026

View reviewed changes

HouMinXi force-pushed the perf/mcp-build-speed-and-memory-rss branch from 28f7bb7 to 99fd19c Compare November 5, 2026 18:25

HouMinXi force-pushed the perf/mcp-build-speed-and-memory-rss branch from 99fd19c to 23c1021 Compare November 6, 2026 12:32

HouMinXi added 2 commits November 7, 2026 11:12

perf: accelerate MCP graph build/incremental updates and reduce resid…

9c0f5de

…ent memory (RSS)

test: resolve bandit B110 and add direct unit test coverage for optim…

20eae3f

…ized symbols

HouMinXi force-pushed the perf/mcp-build-speed-and-memory-rss branch from 23c1021 to 3fca15a Compare November 7, 2026 16:12

HouMinXi force-pushed the perf/mcp-build-speed-and-memory-rss branch from 3fca15a to ae44bd2 Compare November 7, 2026 16:59

tirth8205 added a commit that referenced this pull request Nov 9, 2026

fix: add a CHANGELOG entry for the build and memory work (review of #…

0695b31

…1139) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

tirth8205 mentioned this pull request Nov 9, 2026

Merge train: 43 reviewed PRs (2026-11-09) #1186

Merged

tirth8205 closed this pull request by merging all changes into tirth8205:staging in d01c1fd Nov 9, 2026

Navigation Menu

Repository navigation

perf: accelerate MCP graph build/incremental updates and reduce resident memory (RSS) - #1139

perf: accelerate MCP graph build/incremental updates and reduce resident memory (RSS)#1139
3 commits merged into
tirth8205:stagingfrom
HouMinXi:perf/mcp-build-speed-and-memory-rss

HouMinXi commented Oct 28, 2026

Uh oh!

github-actions Bot commented Oct 28, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

HouMinXi Nov 5, 2026

Uh oh!

HouMinXi Nov 5, 2026

Uh oh!

HouMinXi Nov 5, 2026 •

edited

Loading

Uh oh!

HouMinXi Nov 5, 2026

Uh oh!

HouMinXi Nov 5, 2026

Uh oh!

HouMinXi Nov 5, 2026

Uh oh!

HouMinXi commented Nov 5, 2026

Uh oh!

HouMinXi commented Nov 6, 2026

Uh oh!

HouMinXi commented Nov 7, 2026

Uh oh!

HouMinXi commented Nov 7, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

HouMinXi commented Oct 28, 2026

Summary

Key Optimizations

1. Graph Build & Incremental Update Speed (incremental.py, tools/build.py, communities.py, flows.py, graph.py, parser.py)

2. Resident Memory (RSS) Reduction (main.py, tools/__init__.py, graph.py, embeddings.py, parser.py, constants.py, daemon.py)

Benchmark Results (378 files, 8,869 nodes, 81,885 edges)

Build & Incremental Speed

Resident & Peak Memory (VmRSS / VmHWM)

Verification & Quality Gates

Uh oh!

github-actions Bot commented Oct 28, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

code-review-graph review

Risk-scored changes

Affected execution flows

Test gaps

Reached only through a caller

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟡 Changes recommended

Uh oh!

HouMinXi Nov 5, 2026

Choose a reason for hiding this comment

Uh oh!

HouMinXi Nov 5, 2026

Choose a reason for hiding this comment

Uh oh!

HouMinXi Nov 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

HouMinXi Nov 5, 2026

Choose a reason for hiding this comment

Uh oh!

HouMinXi Nov 5, 2026

Choose a reason for hiding this comment

Uh oh!

HouMinXi Nov 5, 2026

Choose a reason for hiding this comment

Uh oh!

HouMinXi commented Nov 5, 2026

Uh oh!

HouMinXi commented Nov 6, 2026

Uh oh!

HouMinXi commented Nov 7, 2026

Uh oh!

HouMinXi commented Nov 7, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

1. Graph Build & Incremental Update Speed (`incremental.py`, `tools/build.py`, `communities.py`, `flows.py`, `graph.py`, `parser.py`)

2. Resident Memory (RSS) Reduction (`main.py`, `tools/init.py`, `graph.py`, `embeddings.py`, `parser.py`, `constants.py`, `daemon.py`)

Resident & Peak Memory (`VmRSS` / `VmHWM`)

github-actions Bot commented Oct 28, 2026 •

edited

Loading

HouMinXi Nov 5, 2026 •

edited

Loading