Performance

Crate suite

Across 3,405 package versions, the unweighted means are 0.162 s elapsed and 30.0 MiB peak RSS.

Each point averages three scans of one package version. Axes are logarithmic. dioxus-free-icons 0.10.0 averages 36.44 s and 7,324.7 MiB.

Red triangles highlight tokio 1.53.3, and black diamonds highlight diesel 2.3.14.

Methodology

Scans use dejadoc at c11dfd804 with --functions --context-blocks --all-targets --json --no-fail on two logical CPUs of an AMD Threadripper PRO 5975WX. Runs are serial, with nice -n 10, an isolated CARGO_HOME and a populated offline registry.

Each batch is scanned in three permuted passes without clearing the operating-system cache.

LOC counts physical lines in packaged .rs files, including comments, tests and examples, excluding target, .git and symlinks. Some files lie outside scanned targets.

Elapsed time includes process launch, timing wrappers and cargo metadata, excluding compilation and installation. Peak RSS averages GNU time measurements across three runs without summing process-tree memory.

Findings

Each point is one canonical duplicate group, measured in weighted tokens per copy and occurrences. Both axes are logarithmic.

Functions and contexts have a 30-token floor. The 23 word-count fallback doctest groups are excluded. Findings from different checks can overlap.

Diesel case study

Measured on Diesel at a429c50c8 with the crate-suite binary and host.

Elapsed times are medians of three runs, except for single empty-cache runs. Peak RSS is the largest GNU time measurement. Cargo uses an isolated CARGO_HOME, stable Rust and a populated offline registry, without clearing the filesystem cache.

Local checks

All configurations include doctests. The full scan covers 537 doctests, 5,503 functions and 12,354 contexts.

Configuration	Elapsed	Peak RSS
Default	0.76 s	196 MiB
--functions	1.66 s	257 MiB
--context-blocks	1.88 s	234 MiB
Both syntax checks	2.77 s	257 MiB
Both plus --all-targets	2.78 s	257 MiB
Both, restricted with -p diesel	1.68 s	184 MiB
Both, explicit workspace context selection and search	2.93 s	372 MiB
Both, select transaction_manager.rs against the workspace	2.92 s	372 MiB
Both, function and context minimum sizes set to 100	2.77 s	257 MiB

Explicit scopes use --all-targets, selecting either workspace or file:diesel/src/connection/transaction_manager.rs against workspace.

-p limits discovery unless explicit context scopes require more targets. Those scopes filter a workspace index and retain a copied target set, so file and workspace selections can have similar costs.

Token floors apply after canonicalization, reducing reports without avoiding analysis. --all-targets adds work when a workspace has additional binaries or examples.

Dependency exploration

All rows use --functions --context-blocks --all-targets. Only the final row enables latest-version checks. Empty-cache runs start with a fresh dejadoc cache, which warm runs reuse.

Configuration	Reached packages	Empty cache	Warm cache	Peak RSS	Coverage
Depth 1	87	10.22 s	6.53 s	516 MiB	Complete
Depth 2	155	19.26 s	9.76 s	694 MiB	Partial
Depth 3	186	23.99 s	13.43 s	1,572 MiB	Partial
Depth 5	213	24.50 s	14.19 s	1,592 MiB	Partial
Depth 1, include development and build dependencies	92	10.39 s	6.28 s	533 MiB	Complete
Depth 1, all Cargo features	90	10.69 s	6.29 s	525 MiB	Complete
Depth 1, explore all optional dependencies	90	11.50 s	7.00 s	526 MiB	Complete
Depth 1, latest-version checks enabled	87	Not measured	30.44 s	527 MiB	Complete

Depths 2, 3 and 5 exit with 2 and report reuse.coverage.complete = false. The scanner cannot resolve the build-script-selected rustix backend. Warm runs rebuild the failed package, so these rows measure partial analyses.

Latest-version and source-provenance checks can require HTTP even with warm indexes. Cargo's offline mode does not control dejadoc's HTTP client.

See Dependency reuse for depth, feature, cache and coverage semantics.

Scaling limits

• Discovery retains source text and syntax trees, including doctest-only scans.
• Dependency traversal visits each package once.
• Dependency records and canonical strings remain in memory. Warm indexes still require parsing, hashing and deserialization.
• Common canonical forms can produce many candidate pairs, and containment suppression compares duplicate groups.
• Canonical analysis caps regions at 16,384 nodes, nesting at 64 levels and expanded search at 4,096 states.
• The scan pipeline is largely serial.

There is no whole-scan memory or time budget. Use -p and --context-check-latest false to reduce work.

Measure a workspace

Run from the workspace root on Linux with CPU IDs 0,1 available and a dependency-exploration build.

cache=$(mktemp -d)
taskset -c 0,1 nice -n 10 /usr/bin/time \
  -f '%e s elapsed, %M KiB peak RSS' \
  cargo-dejadoc dejadoc --functions --context-blocks --all-targets \
  --context-dependency-depth 1 --context-check-latest false \
  --context-cache-dir "$cache" --json --no-fail > report.json

Reuse the cache for warm runs and create a fresh one for each cold configuration. Check reuse.coverage.complete, indexed-file counts and exit status. Record the tool version, source revision, features and network settings.