adityamukho · Oct 30, 2026 · Oct 30, 2026 · Oct 30, 2026 · Oct 30, 2026
@@ -11,7 +11,7 @@ permissions:
 
 jobs:
   binary-size:
-    name: Check release binary size
+    name: Check distribution binary size
     runs-on: ubuntu-latest
     env:
       SIZE_LIMIT_BYTES: 1263616  # 1234 KiB — #277 error codes, #306 top-k, #723 ancestor join planning; see PHILOSOPHY.md §4
@@ -22,12 +22,13 @@ jobs:
       - name: Install Rust toolchain
         uses: dtolnay/rust-toolchain@stable
 
-      - name: Build release binary
-        run: cargo build --release
+      # The profile cargo-dist ships (#796)
+      - name: Build distribution binary
+        run: cargo build --profile dist
 
       - name: Check binary size
         run: |
-          BINARY=target/release/minigraf
+          BINARY=target/dist/minigraf
           SIZE=$(stat -c %s "$BINARY")
           SIZE_KB=$(( SIZE / 1024 ))
           LIMIT_KB=$(( SIZE_LIMIT_BYTES / 1024 ))

@@ -26,6 +26,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Performance
 
+- **Released binaries use full LTO** (#796). The `dist` profile cargo-dist builds releases with no longer overrides `release`'s full LTO with thin LTO. The CLI is 7–31% smaller on the five release targets (x86_64 Linux 1,501,232 → 1,262,584 bytes, measured before #794) and 4–17% faster on query workloads; clean builds take up to 10 seconds longer. `binary-size.yml` now builds `--profile dist`, so it measures the shipped x86_64 Linux binary (still under the unchanged 1,263,616-byte cap). Windows packages keep cargo-dist's static C runtime. See `docs/benchmarks/796-dist-lto.md`.
+
 - **Sort set element encodings without a stable sort** (#794). `set_elements` sorts its elements' byte encodings with `sort_unstable`: equal encodings are identical bytes and deduplicated straight after, so the encoded bytes are unchanged, and the stable sort's scratch allocation and code go. The native release binary is 3,024 bytes smaller (WASI 950, thin-LTO dist 4,112). On the bench machine, ingest, checkpoint, set-constant queries and scans of 100,000 32-element sets in seven input shapes stayed within run-to-run noise. No API, format or error change.
 
 - **Unstable sorts for unique query metadata** (#795). Prepared-query slot names, an `or`'s branch variables and a subplan's imported variables sort with `sort_unstable`: their keys are unique, or, for a join variable named twice, the tied entries are identical, so order and plans are unchanged. Release builds shrink by 1,008 bytes (native CLI), 1,752 bytes (WASI) and 1,886 bytes (browser `--lib`); prepare and planning time is unchanged within noise. No API, dependency or file-format change.

@@ -130,10 +130,10 @@ lto = true
 codegen-units = 1
 panic = "unwind"
 
-# The profile that 'dist' will build with
+# The profile that 'dist' will build with. It keeps release's full LTO, so the
+# shipped binaries are the ones binary-size.yml measures (#796).
 [profile.dist]
 inherits = "release"
-lto = "thin"
 
 [lints]
 workspace = true

@@ -118,7 +118,12 @@ SQLite's success comes from a clear philosophy: be a library, not a server. Be s
   fit the 2 KiB left after the streaming engine (#473) and set values (#316).
   The approved #723 ancestor-join optimization raises it to 1,234 KiB:
   the measured increase is 2,720 bytes, avoiding a whole-attribute value
-  map for selective in-memory recursion without adding a storage index
+  map for selective in-memory recursion without adding a storage index.
+  The gate builds the x86_64 Linux CLI with cargo-dist's `dist` profile,
+  the one releases ship. From v3.0.0 `dist` keeps release's full LTO
+  rather than thin LTO, 7–31% smaller per target (#796; sizes per target
+  in `docs/benchmarks/796-dist-lto.md`). Other targets are measured, not
+  gated; a profile setting for one target needs its own measurement
 - No runtime dependencies (no JVM, no Python, no Node.js)
 
 **Anti-pattern**: Requiring external services, libraries, or runtimes to function.

@@ -0,0 +1,100 @@
+# Full LTO for distribution builds (#796)
+
+cargo-dist builds the released CLI with the `dist` profile. Before #796 it
+inherited `release` but set `lto = "thin"`, while `binary-size.yml` built
+`--release` (full LTO): the gate measured a binary nobody downloaded. `dist`
+now keeps release's full LTO, and the gate builds `--profile dist`.
+
+## Method
+
+Revision: v3 `2ff91059e` (before #794) plus this change's profile edit, as
+    temporary commits `36598e361` (matrix and bench) and `ac7e661c2` (Windows
+static-CRT check) that also added the measuring workflow; rustc 1.99.1
+everywhere. Thin LTO is the same profile with
+`CARGO_PROFILE_DIST_LTO=thin`.
+
+- **CI matrix**: the five cargo-dist targets on the runners `dist plan` assigns
+  (macos-14, ubuntu-22.04-arm, macos-15-intel, windows-2022, ubuntu-22.04),
+  run once from a temporary branch (`measure/796`, workflow run 41012917889;
+  the branch is deleted). Per runner: [`builds.py`](796-dist-lto/builds.py)
+  times four clean builds (thin, full, full, thin; dependencies fetched, not
+  compiled), [`runtime.py`](796-dist-lto/runtime.py) runs 10 alternating
+  rounds, and `dist build --artifacts=local` packages the target, whose
+  archive is unpacked and run on `demos/demo_commands.txt`.
+- **Bench machine**: the same two scripts on x86_64 Linux, bound to one NUMA
+  node, with 20 runtime rounds.
+- **Runtime workloads** (each a CLI process fed a script on stdin; times include
+  start-up, ingestion and printing): `count` (#772's 2,000 facts, 400 counts),
+  `recursive` (closure of a 300-node chain, 10 queries), `aggregate` (5,000
+  people in 20 departments, a join and grouped sum, 50 queries), `file` (20
+  transactions of 500 facts into a `.graph` file, one count). Both builds must
+  print the same output, transaction timestamps aside.
+
+## Size
+
+| Target | Thin LTO | Full LTO | Saved |
+| --- | ---: | ---: | ---: |
+| x86_64-unknown-linux-gnu | 1,501,232 | **1,262,584** | 238,648 (15.9%) |
+| aarch64-unknown-linux-gnu | 1,262,544 | 1,094,504 | 168,040 (13.3%) |
+| x86_64-apple-darwin | 1,503,392 | 1,075,840 | 427,552 (28.4%) |
+| aarch64-apple-darwin | 1,372,560 | 954,192 | 418,368 (30.5%) |
+| x86_64-pc-windows-msvc, as packaged (static CRT) | 1,139,712 | 1,060,864 | 78,848 (6.9%) |
+
+The x86_64 Linux full-LTO binary is what the gate measures: 1,262,584 bytes
+against the unchanged 1,263,616-byte cap. It equals the `--release` binary, so
+the change gives the gate no headroom. Sizes were identical across a runner's
+two builds.
+
+**Packaging.** On Linux and macOS the packaged binary has the full-LTO size
+(aarch64-apple-darwin: 16 bytes more, the build is under a different target
+directory); its bytes differ from the separate build. On Windows cargo-dist
+links the C runtime statically (`+crt-static`), which the plain builds above
+do not: a second run (41014187879) built thin and full LTO with
+`RUSTFLAGS=-Ctarget-feature=+crt-static` and the `--target` cargo-dist uses.
+The full build matches the packaged 1,060,864 bytes, differing only in the PE
+header; thin is 1,139,712
+([`package-x86_64-pc-windows-msvc-static-crt.json`](796-dist-lto/package-x86_64-pc-windows-msvc-static-crt.json)). Without the static CRT the Windows binaries are
+1,041,920 (thin) and 963,072 (full). All five packaged binaries ran the demo.
+
+The gate checks x86_64 Linux only. The other targets' sizes are recorded here,
+not gated.
+
+## Build time
+
+Clean builds, seconds, in run order (thin, full, full, thin):
+
+| Runner | Thin | Full | Full | Thin |
+| --- | ---: | ---: | ---: | ---: |
+| ubuntu-22.04 | 43 | 30 | 31 | 28 |
+| ubuntu-22.04-arm | 30 | 28 | 28 | 26 |
+| macos-15-intel | 124 | 97 | 77 | 67 |
+| macos-14 | 42 | 41 | 42 | 37 |
+| windows-2022 | 64 | 46 | 47 | 46 |
+| bench, x86_64 Linux | 28 | 31 | 31 | 28 |
+
+The first build on each runner is slower whatever its profile (cold caches).
+Comparing the warmer pairs, full LTO costs 0–10 seconds (0–16%) per build,
+3 seconds on the bench machine. A release builds each target once.
+
+## Runtime
+
+Median full-vs-thin change in elapsed time, and the rounds full LTO won:
+
+| Runner | count | recursive | aggregate | file |
+| --- | ---: | ---: | ---: | ---: |
+| ubuntu-22.04 | −12.7% (10/10) | −9.4% (10/10) | −8.7% (10/10) | −9.0% (10/10) |
+| ubuntu-22.04-arm | −9.7% (10/10) | −16.0% (10/10) | −17.4% (10/10) | −3.6% (10/10) |
+| macos-15-intel | −7.4% (7/10) | −3.9% (9/10) | −7.8% (8/10) | −1.5% (6/10) |
+| macos-14 | −6.3% (8/10) | −10.3% (10/10) | −5.7% (8/10) | +3.5% (5/10) |
+| windows-2022 | −5.7% (10/10) | −9.4% (10/10) | −6.7% (10/10) | −2.1% (7/10) |
+| bench, x86_64 Linux | −9.5% (20/20) | −8.4% (20/20) | −11.0% (20/20) | −4.3% (13/20) |
+
+Full LTO is faster on every query-bound workload on every target. `file`
+(about 0.1–0.3 s, mostly process start-up and file I/O) is within noise on
+macOS. Hosted runners are shared machines; the bench rows are the controlled
+measurement.
+
+## Evidence
+
+[`796-dist-lto/`](796-dist-lto/): `builds-*.json`, `runtime-*.json` (every
+round) and `package-*.json` per target, plus the two scripts.
@@ -0,0 +1,30 @@
+{
+  "platform": "macOS-14.8.9-arm64-arm-64bit",
+  "machine": "arm64",
+  "builds": [
+    {
+      "round": 0,
+      "lto": "thin",
+      "build_seconds": 42.01392458400005,
+      "bytes": 1372560
+    },
+    {
+      "round": 0,
+      "lto": "full",
+      "build_seconds": 40.54684737499997,
+      "bytes": 954192
+    },
+    {
+      "round": 1,
+      "lto": "full",
+      "build_seconds": 42.17448854100002,
+      "bytes": 954192
+    },
+    {
+      "round": 1,
+      "lto": "thin",
+      "build_seconds": 36.50212337500005,
+      "bytes": 1372560
+    }
+  ]
+}
@@ -0,0 +1,30 @@
+{
+  "platform": "Linux-6.8.0-1070-azure-aarch64-with-glibc2.35",
+  "machine": "aarch64",
+  "builds": [
+    {
+      "round": 0,
+      "lto": "thin",
+      "build_seconds": 30.349435244999995,
+      "bytes": 1262544
+    },
+    {
+      "round": 0,
+      "lto": "full",
+      "build_seconds": 28.18664725100001,
+      "bytes": 1094504
+    },
+    {
+      "round": 1,
+      "lto": "full",
+      "build_seconds": 28.104224689000034,
+      "bytes": 1094504
+    },
+    {
+      "round": 1,
+      "lto": "thin",
+      "build_seconds": 25.75162138500002,
+      "bytes": 1262544
+    }
+  ]
+}
@@ -0,0 +1,30 @@
+{
+  "platform": "Linux-6.8.0-139-generic-x86_64-with-glibc2.39",
+  "machine": "x86_64",
+  "builds": [
+    {
+      "round": 0,
+      "lto": "thin",
+      "build_seconds": 27.877472235122696,
+      "bytes": 1501232
+    },
+    {
+      "round": 0,
+      "lto": "full",
+      "build_seconds": 30.649512137984857,
+      "bytes": 1262584
+    },
+    {
+      "round": 1,
+      "lto": "full",
+      "build_seconds": 30.725669295992702,
+      "bytes": 1262584
+    },
+    {
+      "round": 1,
+      "lto": "thin",
+      "build_seconds": 27.925453558098525,
+      "bytes": 1501232
+    }
+  ]
+}
@@ -0,0 +1,30 @@
+{
+  "platform": "macOS-15.7.9-x86_64-i386-64bit",
+  "machine": "x86_64",
+  "builds": [
+    {
+      "round": 0,
+      "lto": "thin",
+      "build_seconds": 123.61216050499996,
+      "bytes": 1503392
+    },
+    {
+      "round": 0,
+      "lto": "full",
+      "build_seconds": 96.794796571,
+      "bytes": 1075840
+    },
+    {
+      "round": 1,
+      "lto": "full",
+      "build_seconds": 76.51469797699997,
+      "bytes": 1075840
+    },
+    {
+      "round": 1,
+      "lto": "thin",
+      "build_seconds": 67.28196448799997,
+      "bytes": 1503392
+    }
+  ]
+}
@@ -0,0 +1,30 @@
+{
+  "platform": "Windows-2022Server-10.0.20348-SP0",
+  "machine": "AMD64",
+  "builds": [
+    {
+      "round": 0,
+      "lto": "thin",
+      "build_seconds": 63.82672539999999,
+      "bytes": 1041920
+    },
+    {
+      "round": 0,
+      "lto": "full",
+      "build_seconds": 46.1428391,
+      "bytes": 963072
+    },
+    {
+      "round": 1,
+      "lto": "full",
+      "build_seconds": 47.27537620000004,
+      "bytes": 963072
+    },
+    {
+      "round": 1,
+      "lto": "thin",
+      "build_seconds": 45.59298579999995,
+      "bytes": 1041920
+    }
+  ]
+}
@@ -0,0 +1,30 @@
+{
+  "platform": "Linux-6.8.0-1070-azure-x86_64-with-glibc2.35",
+  "machine": "x86_64",
+  "builds": [
+    {
+      "round": 0,
+      "lto": "thin",
+      "build_seconds": 43.43511368700001,
+      "bytes": 1501232
+    },
+    {
+      "round": 0,
+      "lto": "full",
+      "build_seconds": 30.330564745999993,
+      "bytes": 1262584
+    },
+    {
+      "round": 1,
+      "lto": "full",
+      "build_seconds": 30.577716478000013,
+      "bytes": 1262584
+    },
+    {
+      "round": 1,
+      "lto": "thin",
+      "build_seconds": 28.490207176000013,
+      "bytes": 1501232
+    }
+  ]
+}
@@ -0,0 +1,26 @@
+#!/usr/bin/env python3
+"""Time clean thin- and full-LTO `dist` builds, two rounds, alternating (#796).
+
+Usage: builds.py OUTPUT_JSON  (from the repository root; copies each binary
+to bins/thin-minigraf and bins/full-minigraf for runtime.py)
+"""
+import json, os, platform, shutil, subprocess, sys, time
+exe = "minigraf.exe" if os.name == "nt" else "minigraf"
+rows = []
+for rnd, order in enumerate([("thin", "full"), ("full", "thin")]):
+    for label in order:
+        tdir = f"tgt-{label}-{rnd}"
+        env = dict(os.environ)
+        if label == "thin":
+            env["CARGO_PROFILE_DIST_LTO"] = "thin"
+        start = time.perf_counter()
+        subprocess.run(["cargo", "build", "--locked", "--profile", "dist", "--target-dir", tdir], env=env, check=True)
+        secs = time.perf_counter() - start
+        size = os.path.getsize(os.path.join(tdir, "dist", exe))
+        rows.append({"round": rnd, "lto": label, "build_seconds": secs, "bytes": size})
+        print(rows[-1], flush=True)
+        os.makedirs("bins", exist_ok=True)
+        shutil.copy(os.path.join(tdir, "dist", exe), os.path.join("bins", f"{label}-{exe}"))
+        shutil.rmtree(tdir)
+json.dump({"platform": platform.platform(), "machine": platform.machine(), "builds": rows},
+          open(sys.argv[1], "w"), indent=2)
@@ -0,0 +1,7 @@
+{
+  "target": "aarch64-apple-darwin",
+  "packaged_bytes": 954208,
+  "full_bytes": 954192,
+  "thin_bytes": 1372560,
+  "packaged_equals_full": false
+}
@@ -0,0 +1,7 @@
+{
+  "target": "aarch64-unknown-linux-gnu",
+  "packaged_bytes": 1094504,
+  "full_bytes": 1094504,
+  "thin_bytes": 1262544,
+  "packaged_equals_full": false
+}