fix(build): recompute a stale package snapshot, not just a missing one (#190)

## Motivation

`arm64/alpine324` reported nothing to build while thousands were missing:

```
line  49: Precomputed 7192 package versions (6871 r-minor-sensitive)   <- install-deps agent
line  99: Total# of remaining package versions: 43 (sensitive_only=TRUE) <- a build shard
line 101: Skipped 0 package versions already attempted under R 4.4; 0 remaining
```

Both numbers come from the **same pipeline**. The same run's index step dropped 2407 packages as missing for 4.4 and 2436 for 4.6.

## Cause

```r
if (!all(file.exists(package_cache_files))) { ... recompute ... }
```

Existence is not freshness. The snapshot describes S3 and CRAN state when it was written, and the cache volume is per-agent — the file's own comment says so. An agent that ran an earlier pipeline keeps serving that pipeline's answer forever, and no later fix to how the snapshot is computed (#189) can reach it.

## Change

Recompute when the snapshot is stale as well as when it is missing. Keyed on the pipeline when the CI exposes an identifier (`CI_PIPELINE_NUMBER`, `CI_BUILD_NUMBER`, `CI_PIPELINE_ID`), so a new pipeline recomputes once per agent and its shards then share the result. Off CI, or when none is set, an age check with a two hour default (`PACKAGE_SNAPSHOT_TTL_HOURS`).

## Verification

| scenario | decision |
|---|---|
| files missing | RECOMPUTE |
| same pipeline id | reuse |
| **new pipeline id** | **RECOMPUTE** |
| no CI var, recent file | reuse |
| no CI var, aged out | RECOMPUTE |

I could not confirm which identifier Crow actually sets — none is referenced anywhere in this repo — so all three are tried and the age check backs them up. If none is present the behaviour is the age path, which is still correct, just coarser.

Reviewed-on: #190

.crow/build-all-versions.yaml

@@ -35,6 +35,14 @@ variables:
       - 'resolute'
     default: '3.24'
   R_VERSION:
+    # The slot's *primary* R minor: what `local/build-all.R` builds into the
+    # generic slot. The loop below already runs `--sensitive-only` for every
+    # other installed minor, so filling a non-primary minor's gap needs this
+    # left alone, not changed.
+    #
+    # 4.6.0 was briefly offered here (#183) and removed: selecting it for a
+    # slot whose generic binaries are 4.5-built would publish 4.6 binaries
+    # into the generic slot and break every 4.5 client.
     description: 'Primary R version under /opt/R.'
     options:
       - 4.5.3

.crow/reindex.yaml

@@ -172,7 +172,12 @@ steps:
         from_secret: BUNNYNET_API_KEY
       # cran.rpkgs.com and cran.allianceswisspass.devxy.io are on separate
       # Bunny pull zones, so both must be purged after the shared origin changes.
-      BUNNY_PULLZONES: '3857050 cran.allianceswisspass.devxy.io'
+      # The staging zone is listed too. It shares the B2 origin, so an index
+      # it still holds is a stale copy of the same object, and its
+      # cache_expiration_time is the same ~370 days: without a purge here it
+      # serves pre-reindex indexes indefinitely and any verification run
+      # against it measures the past.
+      BUNNY_PULLZONES: '3857050 cran.allianceswisspass.devxy.io cran-rpkgs-test.b-cdn.net'
     commands:
       - apk add --no-cache -q bash curl jq
       # Crow carries the checkout from the re-index step into this step.

.crow/weekly-rebuild-reindex.yaml

@@ -175,7 +175,12 @@ steps:
         from_secret: BUNNYNET_API_KEY
       # cran.rpkgs.com and cran.allianceswisspass.devxy.io are on separate
       # Bunny pull zones, so both must be purged after the shared origin changes.
-      BUNNY_PULLZONES: '3857050 cran.allianceswisspass.devxy.io'
+      # The staging zone is listed too. It shares the B2 origin, so an index
+      # it still holds is a stale copy of the same object, and its
+      # cache_expiration_time is the same ~370 days: without a purge here it
+      # serves pre-reindex indexes indefinitely and any verification run
+      # against it measures the past.
+      BUNNY_PULLZONES: '3857050 cran.allianceswisspass.devxy.io cran-rpkgs-test.b-cdn.net'
     commands:
       - apk add --no-cache -q bash curl jq
       # Crow carries the checkout from the re-index step into this step.

cdn.tf

@@ -91,7 +91,17 @@ resource "bunnynet_compute_script" "rpkgs_router" {
 resource "bunnynet_compute_script_variable" "rpkgs_router_union_slots" {
   script = bunnynet_compute_script.rpkgs_router.id
   name   = "UNION_SLOTS"
-  default_value = ""
+  # Enabled. Until this was set, every client resolved against the generic
+  # index and never reached a per-minor binary: an R 4.6.1 client on resolute
+  # downloaded the 4.5-built rlang (2079570 bytes) while the correct 4.6 build
+  # (2075106 bytes) sat unused one directory away, and died at load with
+  # `undefined symbol: SETLENGTH`.
+  #
+  # Verified before enabling, against the staging zone with the same script and
+  # the same origin: all 16 slots report zero regressions against the generic
+  # slot, an excluded R minor is sent to CRAN, a client without an R minor
+  # still gets the generic index, and tarball requests are never rewritten.
+  default_value = join(",", local.rpkgs_slots)
   required      = false
 }
 

local/build-all.R

@@ -26,9 +26,50 @@ package_cache_files <- c(
   "/mnt/cache/packages/r_minor_sensitive_pkgs.rds",
   "/mnt/cache/packages/s3_cache.rds"
 )
+# Existence is not freshness. The snapshot describes S3 and CRAN state at the
+# moment it was written, and the volume is per-agent, so an agent that ran an
+# earlier pipeline keeps serving that pipeline's answer forever. arm64/alpine324
+# reported "0 remaining" for both 4.4 and 4.6 from a stale snapshot listing 43
+# sensitive packages, while the install-deps step in the very same pipeline had
+# just computed 6871 on another agent.
+#
+# Keyed on the pipeline when the CI exposes one, so a new pipeline recomputes
+# once per agent and its shards then share the result. Off CI, or when no such
+# variable is set, fall back to an age check.
+snapshot_id_path <- "/mnt/cache/packages/snapshot.id"
+snapshot_ttl_hours <- as.numeric(
+  Sys.getenv("PACKAGE_SNAPSHOT_TTL_HOURS", unset = "2")
+)
+current_snapshot_id <- ""
+for (v in c("CI_PIPELINE_NUMBER", "CI_BUILD_NUMBER", "CI_PIPELINE_ID")) {
+  val <- Sys.getenv(v, unset = "")
+  if (nzchar(val)) {
+    current_snapshot_id <- paste(v, val, sep = "=")
+    break
+  }
+}
+
+snapshot_is_stale <- function() {
   if (!all(file.exists(package_cache_files))) {
+    return(TRUE)
+  }
+  if (nzchar(current_snapshot_id)) {
+    cached <- tryCatch(
+      readLines(snapshot_id_path, warn = FALSE)[1L],
+      error = function(e) NA_character_,
+      warning = function(w) NA_character_
+    )
+    return(!identical(cached, current_snapshot_id))
+  }
+  age_hours <- as.numeric(
+    difftime(Sys.time(), file.mtime(package_cache_files[1L]), units = "hours")
+  )
+  isTRUE(age_hours > snapshot_ttl_hours)
+}
+
+if (snapshot_is_stale()) {
   message(
-    "Package snapshot missing from cache; recomputing via packages-to-build.R"
+    "Package snapshot missing or stale; recomputing via packages-to-build.R"
   )
   dir.create("/mnt/cache/packages", showWarnings = FALSE, recursive = TRUE)
   save_rds_atomic <- function(obj, path) {
@@ -42,6 +83,9 @@ if (!all(file.exists(package_cache_files))) {
     pkgs[r_minor_sensitive == TRUE],
     "/mnt/cache/packages/r_minor_sensitive_pkgs.rds"
   )
+  if (nzchar(current_snapshot_id)) {
+    writeLines(current_snapshot_id, snapshot_id_path)
+  }
   message("Package snapshot recomputed.")
 }
 
@@ -110,10 +154,25 @@ con <- DBI::dbConnect(
   password = Sys.getenv("PGPASS"),
   sslmode = "require"
 )
+# Scope the skip to the R minor this pass is running under. `single_builds`
+# records `r_version` per attempt, but querying without it made a non-primary
+# pass skip everything the primary pass had already attempted under a different
+# minor - so `--sensitive-only` under 4.6 skipped packages that had only ever
+# been built for 4.5, and the per-minor slots never filled. That is why
+# amd64/resolute served 4000 fewer packages to a 4.6 client than to a 4.5 one.
+r_minor <- paste(
+  R.version$major,
+  strsplit(R.version$minor, ".", fixed = TRUE)[[1L]][1L],
+  sep = "."
+)
 built <- DBI::dbGetQuery(
   con,
-  "SELECT name, tag FROM single_builds WHERE platform = $1 AND arch = $2",
+  paste(
-  params = list(platform, arch)
+    "SELECT name, tag FROM single_builds",
+    "WHERE platform = $1 AND arch = $2",
+    "AND substring(r_version from '^[0-9]+[.][0-9]+') = $3"
+  ),
+  params = list(platform, arch, r_minor)
 )
 DBI::dbDisconnect(con)
 before <- nrow(chunk)
@@ -121,8 +180,9 @@ chunk <- chunk[
   !paste(chunk$Package, chunk$Version) %in% paste(built$name, built$tag),
 ]
 sprintf(
-  "Skipped %d already-attempted package versions; %d remaining for this job",
+  "Skipped %d package versions already attempted under R %s; %d remaining for this job",
   before - nrow(chunk),
+  r_minor,
   nrow(chunk)
 )
 

local/packages-to-build.R

@@ -68,9 +68,36 @@ archive_versions <- archive_versions[
 ]
 
 # Now get release versions (assuming cran_release has Package and Version columns)
+#
+# Packages published in the last few days are held back. `check_for_binary()`
+# reads the published version from the `cran` GitHub mirror
+# (`GET /repos/cran/<pkg>/commits`), and that mirror lags CRAN: a package that
+# has just appeared has no repository there yet. The call then 404s, which is
+# permanent, but it is wrapped in `purrr::insistently` and retried ten times
+# with a backoff capped at 60s - so one unmirrored package burns about five
+# minutes and then aborts the whole shard.
+#
+# Holding them back costs nothing: the daily update pipeline builds new and
+# updated packages anyway, and they arrive here on the next run once the mirror
+# has caught up.
+mirror_lag_days <- as.numeric(
+  Sys.getenv("CRAN_MIRROR_LAG_DAYS", unset = "3")
+)
+published <- as.POSIXct(cran_release$Published, tz = "UTC")
+too_recent <- !is.na(published) &
+  published > (Sys.time() - mirror_lag_days * 86400)
+if (any(too_recent)) {
+  message(sprintf(
+    "Holding back %d package(s) published in the last %g day(s); the cran GitHub mirror will not have them yet: %s",
+    sum(too_recent),
+    mirror_lag_days,
+    paste(utils::head(cran_release$Package[too_recent], 10L), collapse = ", ")
+  ))
+}
+
 release_versions <- data.table(
-  Package = cran_release$Package,
+  Package = cran_release$Package[!too_recent],
-  Version = as.character(cran_release$Version)
+  Version = as.character(cran_release$Version[!too_recent])
 )
 
 pkgs_to_build <- unique(rbind(archive_versions, release_versions, fill = TRUE))
@@ -89,7 +116,28 @@ s3_pkgs <- s3fs::s3_dir_ls(
   recurse = TRUE
 )
 
-file_names <- basename(s3_pkgs)
+# `recurse = TRUE` walks the per-minor slots as well, and `basename()` throws
+# the directory away - so `4.5/curl_1.0.tar.gz` and `curl_1.0.tar.gz` collapse
+# to one name and a package present under *any* R minor counts as built for
+# *all* of them. The candidate list then prunes exactly the packages a
+# per-minor pass exists to build: arm64/alpine324 reported "0 remaining" for
+# both 4.4 and 4.6 while its indexes were dropping 2400+ packages as missing.
+#
+# Per-minor objects are therefore excluded here. Presence in a specific minor
+# is decided downstream, where the running R version is known: build-all.R
+# filters on it, and `build_binary_package()` checks the per-minor path per
+# package and skips what is already there.
+#
+# Archive/ is kept. Those are versions that were built and then superseded;
+# dropping them would make every archived version look unbuilt.
+per_minor_object <- grepl("/[0-9]+\\.[0-9]+/[^/]+$", s3_pkgs)
+if (any(per_minor_object)) {
+  cat(sprintf(
+    "Excluding %d per-minor object(s) from the presence check; those are decided per pass\n",
+    sum(per_minor_object)
+  ))
+}
+file_names <- basename(s3_pkgs[!per_minor_object])
 
 # An object occupying a key is not proof a binary was built: a package whose
 # build failed has its CRAN source published under exactly that name. Left in
@@ -152,11 +200,22 @@ s3_dt <- data.table(
 
 ### Get all packages with build errors
 
+# Scoped to the R minor this snapshot is computed under. A failure is a fact
+# about one interpreter: without the scope a package that failed under the
+# primary minor is dropped from the candidate list for every other minor too,
+# which is the same omission fixed in local/build-all.R and in bincraft's
+# check_package_error().
+snapshot_r_minor <- paste(
+  R.version$major,
+  strsplit(R.version$minor, ".", fixed = TRUE)[[1L]][1L],
+  sep = "."
+)
 sql_query <- paste0(
   # nolint
   "SELECT error_occurred FROM ",
   "single_builds",
-  " WHERE name = $1 AND tag = $2 AND platform = $3 AND arch = $4"
+  " WHERE name = $1 AND tag = $2 AND platform = $3 AND arch = $4",
+  " AND substring(r_version from '^[0-9]+[.][0-9]+') = $5"
 )
 # Function to query for a single package-version
 query_error <- function(pkg, ver) {
@@ -164,7 +223,7 @@ query_error <- function(pkg, ver) {
     ~ DBI::dbGetQuery(
       con,
       sql_query,
-      params = list(pkg, ver, platform, arch)
+      params = list(pkg, ver, platform, arch, snapshot_r_minor)
     ),
     rate = purrr::rate_backoff(
       pause_base = 1L,

scripts/verify-r-minor-routing.sh

@@ -106,10 +106,13 @@ fetch() {
   local status
   # -L: the router answers an index request with a redirect, so the bytes a
   # client ends up with are only visible by following it.
+  #
+  # no-cache: a purge is asynchronous, so a run started right after a reindex
+  # otherwise measures whatever the edge still holds.
   if [ -n "$ua" ]; then
-    status=$(curl -sSL -A "$ua" -o "$dest" -w '%{http_code}' --max-time 120 "$url" 2>/dev/null)
+    status=$(curl -sSL -A "$ua" -H 'Cache-Control: no-cache' -o "$dest" -w '%{http_code}' --max-time 120 "$url" 2>/dev/null)
   else
-    status=$(curl -sSL -o "$dest" -w '%{http_code}' --max-time 120 "$url" 2>/dev/null)
+    status=$(curl -sSL -H 'Cache-Control: no-cache' -o "$dest" -w '%{http_code}' --max-time 120 "$url" 2>/dev/null)
   fi
   echo "$status" > "$dest.status"
   echo "$status"
@@ -137,6 +140,23 @@ fallback_counts() {
   '
 }
 
+# Packages this index serves from the generic slot with a binary built under a
+# different R minor, while some other per-minor slot carries a build of them -
+# which proves the ABI classifier called them risky. Serving those is the
+# load-time crash the per-minor slots exist to prevent. bincraft drops them at
+# index time, so a non-zero count means the slot has not been reindexed since
+# that guard shipped.
+abi_unsafe_count() {
+  local minor_file=$1 minor=$2 risky_file=$3
+  gunzip -c "$minor_file" 2>/dev/null | awk -v m="$minor" '
+    /^Package:/ { pkg = $2; path = ""; built = "" }
+    /^Path:/    { path = $2 }
+    /^Built:/   { built = $2 " " $3 }
+    /^$/        { if (pkg != "" && path == "" && built != "" && built !~ ("^R " m "\\.")) print pkg; pkg = "" }
+  ' | sort -u > "$minor_file.mismatched"
+  comm -12 "$minor_file.mismatched" "$risky_file" | wc -l
+}
+
 # How many packages a client of <minor> would receive as source through
 # per-minor routing while the generic slot holds a binary built under that very
 # minor. Zero is the bar for enabling a slot.
@@ -216,12 +236,25 @@ for arch in $ARCHES; do
       minor_count=$(wc -l < "$minor_file.names")
 
       # Union property: nothing the flat index carries may be missing here.
-      missing=$(comm -23 "$flat_file.names" "$minor_file.names" | head -5)
+      # bincraft deliberately drops an ABI-risky package whose only binary was
-      missing_count=$(comm -23 "$flat_file.names" "$minor_file.names" | wc -l)
+      # built under another R minor: serving it is the load-time crash the
-      if [ "$missing_count" -ne 0 ]; then
+      # per-minor slots exist to prevent. Those absences are correct.
-        bad "$slot R $minor index is not a union: $missing_count flat packages absent (e.g. $(echo "$missing" | tr '\n' ' '))"
+      #
+      # What must never go missing is a generic package built under *this*
+      # minor, which is safe to serve and has no reason to disappear.
+      comm -23 "$flat_file.names" "$minor_file.names" > "$minor_file.absent"
+      absent_count=$(wc -l < "$minor_file.absent")
+      gunzip -c "$flat_file" 2>/dev/null | awk -v m="$minor" '
+        /^Package:/ { pkg = $2; built = "" }
+        /^Built:/   { built = $2 " " $3 }
+        /^$/        { if (pkg != "" && built ~ ("^R " m "\\.")) print pkg; pkg = "" }
+      ' | sort -u > "$minor_file.flatsame"
+      lost=$(comm -12 "$minor_file.absent" "$minor_file.flatsame" | wc -l)
+
+      if [ "${lost:-0}" -ne 0 ]; then
+        bad "$slot R $minor index dropped $lost generic package(s) built under R $minor, which were safe to serve"
       else
-        ok "$slot R $minor index: $minor_count packages, union holds"
+        ok "$slot R $minor index: $minor_count packages, union holds ($absent_count ABI-unsafe dropped)"
       fi
 
       # A per-minor entry that is a source fallback resolves fine but makes the
@@ -311,6 +344,30 @@ for arch in $ARCHES; do
       fi
     fi
 
+    # Packages carrying a Path in any per-minor index are risky by construction.
+    : > "$WORK/${arch}-${distro}.risky"
+    for minor in $MINORS; do
+      f="$WORK/${arch}-${distro}-${minor}.gz"
+      [ -s "$f" ] || continue
+      gunzip -c "$f" 2>/dev/null | awk '
+        /^Package:/ { pkg = $2; path = "" }
+        /^Path:/    { path = $2 }
+        /^$/        { if (pkg != "" && path != "") print pkg; pkg = "" }
+      ' >> "$WORK/${arch}-${distro}.risky"
+    done
+    sort -u -o "$WORK/${arch}-${distro}.risky" "$WORK/${arch}-${distro}.risky"
+
+    for minor in $MINORS; do
+      f="$WORK/${arch}-${distro}-${minor}.gz"
+      [ -s "$f" ] || continue
+      unsafe=$(abi_unsafe_count "$f" "$minor" "$WORK/${arch}-${distro}.risky")
+      if [ "${unsafe:-0}" -gt 0 ]; then
+        bad "$slot R $minor serves $unsafe ABI-risky package(s) built under another R minor - reindex this slot"
+      else
+        ok "$slot R $minor serves no ABI-risky package from another minor"
+      fi
+    done
+
     # Excluded minors: no published index, and under --live a redirect to CRAN.
     for minor in $EXCLUDED_MINORS; do
       ex_url="$BASE/$slot/latest/src/contrib/$minor/PACKAGES.gz"