Give the largest measured per-image memory in "Running under upaas"

The per-image figure is now the largest peak measured over one
8192x8192 source of each source format, with 16-bit PNG, 12-bit AVIF,
and 16-bit and 32-bit floating-point JPEG XL sources too, each saved at
full size to every output format with pixad's libvips settings: 4.2 GiB,
for the 16-bit JPEG XL saved as WebP. Temporary files says that for a
GIF, WebP, AVIF or JPEG XL source the decoded pixels are held in memory
as well as written to /tmp.

Model: opus-5-5

README.md

@@ -130,6 +130,32 @@ What the [upaas](https://git.eeqj.de/sneak/upaas) app for pixa needs:
       the cache can grow to three quarters or more of the space they all have
       left, and to all of it when that is 500 MiB or less.
     - the rest are in the table under Configuration below
+- **Memory:** pixa can use up to `max_concurrent_processing` times 4.2 GiB for
+  the images it processes at once, plus `upstream_connections` times
+  `upstream_max_response_size` for the source images it has fetched that wait
+  for processing. 4.2 GiB is the most resident memory was measured to grow while
+  pixa, with pixad's libvips settings, processed one 8192x8192 source, the
+  largest it accepts, saving it at full size to each output format in turn, one
+  at a time. The sources were a JPEG, a GIF, a WebP, an 8-bit and a 16-bit PNG,
+  an 8-bit and a 12-bit AVIF, and an 8-bit, a 16-bit and a 32-bit floating-point
+  JPEG XL; memory grew most for the 16-bit JPEG XL, of random pixels, saved as
+  WebP. At the defaults that is half the CPUs pixa sees times 4.2 GiB, plus 64
+  times 50 MiB: on a host where pixa sees 48 CPUs, 24 times 4.2 GiB plus 3.1
+  GiB, about 104 GiB. Set `PIXA_MAX_CONCURRENT_PROCESSING` and
+  `PIXA_UPSTREAM_CONNECTIONS` so that this fits the memory the container may
+  use.
+- **Temporary files:** while pixa processes a large source, libvips writes its
+  decoded pixels to a file in `/tmp` and deletes the file when the image is
+  done. For a GIF, WebP, AVIF or JPEG XL source the decoded pixels are held in
+  memory as well as written to that file, and the 4.2 GiB above includes that
+  memory. For an 8192x8192 source the file was measured on pixad at 192 MiB for
+  an 8-bit RGB JPEG, 512 MiB for a 16-bit RGBA PNG or JPEG XL, and 1 GiB for a
+  JPEG XL of 32-bit floating-point RGBA, the largest measured: up to 1 GiB per
+  image processed at once, 24 GiB at the defaults on a host where pixa sees 48
+  CPUs. This `/tmp` is in the container's own filesystem, not on the volume at
+  `/var/lib/pixa`, and `cache_max_bytes` does not count it. The 4.2 GiB above
+  does not include these files; where `/tmp` is a tmpfs, they are memory on top
+  of it.
 - **Health check:** the image's `HEALTHCHECK` requests
   `/.well-known/healthcheck.json`. upaas reads the container's health 60 seconds
   after a deploy and marks the deploy failed unless it is `healthy`. The probe
@@ -164,6 +190,14 @@ name.
 Multiple source paths may reference the same content blob; the database tracks
 references rather than using filesystem refcounting.
 
+Source and transformed image files, and the `.meta` file beside each transformed
+image, are written under a temporary name, synced to disk, and only then renamed
+into place, so a crash leaves no file under its own name with part of its data.
+When pixa starts, and every five minutes after, it removes each cached source or
+transformed image whose file is empty, or not the size the database records for
+it, with a warning naming the file; the next request for it fetches the source
+or converts the image again.
+
 pixa's target is 1-5k r/s, which has not been measured at that rate (see Load
 Test). Toward it, pixa keeps in memory the content types of the 10,000
 transformed images most recently cached or served, so a cache hit on one of them
@@ -545,7 +579,7 @@ tell. With no file, pixa uses the environment and the defaults.
 | `PIXA_ALLOW_HTTP`                    | `allow_http`                    | Allow plain-HTTP upstreams, for testing only; default `false`                |
 | `PIXA_UPSTREAM_CONNECTIONS_PER_HOST` | `upstream_connections_per_host` | Concurrent connections per upstream host; default `20`                       |
 | `PIXA_UPSTREAM_CONNECTIONS`          | `upstream_connections`          | Concurrent connections to all upstream hosts together; default `64`          |
-| `PIXA_MAX_CONCURRENT_PROCESSING`     | `max_concurrent_processing`     | Images processed at once; default the number of CPUs                         |
+| `PIXA_MAX_CONCURRENT_PROCESSING`     | `max_concurrent_processing`     | Images processed at once; default half the CPUs, at least 1                  |
 | `PIXA_IMAGE_REQUESTS_PER_MINUTE`     | `image_requests_per_minute`     | Image requests per client per minute; `0` turns it off; default `600`        |
 | `PIXA_UPSTREAM_FETCH_TIMEOUT`        | `upstream_fetch_timeout`        | Time allowed for one fetch from an upstream host; default `30s`              |
 | `PIXA_UPSTREAM_MAX_RESPONSE_SIZE`    | `upstream_max_response_size`    | Largest upstream response accepted, in bytes; default 50 MiB                 |
@@ -631,10 +665,12 @@ Key settings in more detail:
   does, and `downstream_timeout` has not ended first, the request is answered
   503 with the error `server busy, try again later`
 - `max_concurrent_processing` — the most images decoded and encoded at once;
-  default the number of CPUs pixa can use (`GOMAXPROCS`), which follows a
+  default half the CPUs pixa can use (`GOMAXPROCS`, which follows a container's
-  container's CPU limit. A request that finds all of them in use waits up to 10
+  CPU limit), rounded down, and at least 1. A value that is not a whole number
-  seconds for one to free up; if none does, and `downstream_timeout` has not
+  of at least 1 aborts startup. A request that finds all of them in use waits up
-  ended first, it is answered 503 the same way
+  to 10 seconds for one to free up; if none does, and `downstream_timeout` has
+  not ended first, it is answered 503 the same way. Each one can take a lot of
+  memory: see "Running under upaas"
 - `image_requests_per_minute` — how many requests one client may send to the
   image routes per minute (see the limit on them under Routes); default `600`.
   `0` turns the limit off, and a negative or non-integer value aborts startup.

TODO.md

@@ -30,6 +30,30 @@ P2: security: per-origin rate limiting
 
 # Completed Steps
 
+- 2026-10-15 `max_concurrent_processing` defaults to half the CPUs pixa sees,
+  rounded down, and at least 1 (closes #242): it defaulted to one image per CPU,
+  and a container with no CPU limit sees every CPU of the host, so on a 48-CPU
+  host pixa processed 48 images at once, each of which can take about 4.2 GiB.
+  `config.DefaultMaxConcurrentProcessing` gives the default to the config and to
+  `imageprocessor.New`; `PIXA_MAX_CONCURRENT_PROCESSING` still overrides it.
+  "Running under upaas" in `README.md` gives the worst-case memory, measured,
+  and tells the operator to set `PIXA_MAX_CONCURRENT_PROCESSING` and
+  `PIXA_UPSTREAM_CONNECTIONS` to fit the container's memory. It also gives the
+  file libvips writes to `/tmp` for each large source, up to 1 GiB, measured,
+  which is outside `/var/lib/pixa` and `cache_max_bytes` and is memory where
+  `/tmp` is a tmpfs.
+- 2026-10-15 a cached file left empty or partial is no longer served, or fetched
+  again on every request (closes #252). Source and transformed image files, and
+  each transformed image's `.meta` file, are synced to disk before they are
+  renamed into place, and their directory after; the `.meta` file is written
+  before the image, and a failed directory sync removes the renamed file. The
+  reconciliation pass, at startup and every five minutes, removes a cached
+  source or transformed image whose file is empty, or not the size the database
+  records for it, with its rows and a warning naming the file, so the next
+  request fetches the source or converts the image again; requests compare no
+  sizes. Storing a source over a file of another size replaces that file. A
+  source that fails to decode is not removed for that alone: a damaged upstream
+  image would then be fetched on every request.
 - 2026-10-15 a transcode stops once its request's `downstream_timeout` has
   ended, unless it has begun encoding (closes #243): the image processor looked
   at its context only while it waited for a processing slot, so an image was

configs/config.example.yml

@@ -108,10 +108,11 @@ upstream_connections_per_host: 20
 # downstream_timeout has ended first.
 upstream_connections: 64
 
-# Maximum number of images decoded and encoded at once (default: the
+# Maximum number of images decoded and encoded at once (default: half the
-# number of CPUs pixa can use, which follows a container's CPU limit). A
+# number of CPUs pixa can use, at least 1; the number of CPUs follows a
-# request that finds none free waits up to 10 seconds for one, and if none
+# container's CPU limit). A request that finds none free waits up to 10
-# frees up it is answered 503, unless downstream_timeout has ended first.
+# seconds for one, and if none frees up it is answered 503, unless
+# downstream_timeout has ended first.
 # max_concurrent_processing: 4
 
 # How many requests to the image routes (/v1/image/ and /v1/e/) one client

internal/config/concurrency_limits_internal_test.go

@@ -1,6 +1,7 @@
 package config
 
 import (
+	"fmt"
 	"runtime"
 	"testing"
 )
@@ -12,8 +13,8 @@ const (
 )
 
 // TestOmittedConcurrencyLimitsUseDefaults checks that an omitted
-// max_concurrent_processing is the number of CPUs Go uses and an omitted
+// max_concurrent_processing is half the CPUs Go uses, and at least 1, and
-// upstream_connections is 64.
+// an omitted upstream_connections is 64.
 func TestOmittedConcurrencyLimitsUseDefaults(t *testing.T) {
 	t.Parallel()
 
@@ -22,9 +23,9 @@ func TestOmittedConcurrencyLimitsUseDefaults(t *testing.T) {
 		t.Fatalf("minimal config should be valid, got error: %v", err)
 	}
 
-	if c.MaxConcurrentProcessing != runtime.GOMAXPROCS(0) {
+	if c.MaxConcurrentProcessing != max(runtime.GOMAXPROCS(0)/2, 1) {
-		t.Errorf("MaxConcurrentProcessing = %d, want %d, one per CPU",
+		t.Errorf("MaxConcurrentProcessing = %d, want %d, half the CPUs",
-			c.MaxConcurrentProcessing, runtime.GOMAXPROCS(0))
+			c.MaxConcurrentProcessing, max(runtime.GOMAXPROCS(0)/2, 1))
 	}
 
 	if c.UpstreamConnections != 64 {
@@ -32,6 +33,36 @@ func TestOmittedConcurrencyLimitsUseDefaults(t *testing.T) {
 	}
 }
 
+// TestOmittedMaxConcurrentProcessingIsHalfTheCPUsAndAtLeastOne checks
+// that an omitted max_concurrent_processing is half the CPUs Go uses,
+// rounded down, and 1 when Go uses one CPU.
+//
+//nolint:paralleltest // it sets GOMAXPROCS, which other tests read
+func TestOmittedMaxConcurrentProcessingIsHalfTheCPUsAndAtLeastOne(t *testing.T) {
+	cpusBefore := runtime.GOMAXPROCS(0)
+
+	t.Cleanup(func() { runtime.GOMAXPROCS(cpusBefore) })
+
+	for _, tc := range []struct{ cpus, want int }{
+		{1, 1}, {2, 1}, {3, 1}, {8, 4},
+	} {
+		runtime.GOMAXPROCS(tc.cpus)
+
+		c, err := configFromYAML(t, signingKeyLine)
+		if err != nil {
+			t.Fatalf("minimal config should be valid, got error: %v", err)
+		}
+
+		t.Logf("GOMAXPROCS %d: MaxConcurrentProcessing %d",
+			tc.cpus, c.MaxConcurrentProcessing)
+
+		if c.MaxConcurrentProcessing != tc.want {
+			t.Errorf("GOMAXPROCS %d: MaxConcurrentProcessing = %d, want %d",
+				tc.cpus, c.MaxConcurrentProcessing, tc.want)
+		}
+	}
+}
+
 // TestExplicitConcurrencyLimitsAreUsed checks that valid values for the
 // two limits are used as given.
 func TestExplicitConcurrencyLimitsAreUsed(t *testing.T) {
@@ -160,3 +191,19 @@ func TestConcurrencyLimitsFromEnvironment(t *testing.T) {
 		})
 	}
 }
+
+// TestInvalidMaxConcurrentProcessingFromEnvironmentAbortsStartup checks
+// that a PIXA_MAX_CONCURRENT_PROCESSING that is not a whole number of at
+// least 1, an empty one included, aborts startup naming the variable and
+// the key instead of taking the default.
+func TestInvalidMaxConcurrentProcessingFromEnvironmentAbortsStartup(t *testing.T) {
+	for _, value := range []string{"", " ", "0.5", "half", "-1", "4 images"} {
+		t.Run(fmt.Sprintf("value %q", value), func(t *testing.T) {
+			t.Setenv(testMaxConcurrentProcessingVar, value)
+
+			_, err := configFromYAML(t, signingKeyLine)
+			wantStartupError(t, err,
+				testMaxConcurrentProcessingVar, keyMaxConcurrentProcessing)
+		})
+	}
+}

internal/config/config.go

@@ -37,6 +37,14 @@ const (
 	DefaultImageRequestsPerMinute     = 600
 )
 
+// DefaultMaxConcurrentProcessing returns the most images processed at once
+// when max_concurrent_processing is omitted: half the CPUs Go uses
+// (runtime.GOMAXPROCS(0), which follows a container's CPU limit), rounded
+// down, and at least 1.
+func DefaultMaxConcurrentProcessing() int {
+	return max(runtime.GOMAXPROCS(0)/2, 1) //nolint:mnd // half the CPUs
+}
+
 // Configuration key names.
 const (
 	keyDebug                      = "debug"
@@ -297,10 +305,8 @@ func newFromSmartConfig(sc *smartconfig.Config) (*Config, error) {
 			keyUpstreamConnectionsPerHost, DefaultUpstreamConnectionsPerHost),
 		UpstreamConnections: loader.intVal(
 			keyUpstreamConnections, DefaultUpstreamConnections),
-		// Decoding and encoding are CPU-bound, so the default is one image
-		// per CPU Go uses, which follows a container's CPU limit.
 		MaxConcurrentProcessing: loader.intVal(
-			keyMaxConcurrentProcessing, runtime.GOMAXPROCS(0)),
+			keyMaxConcurrentProcessing, DefaultMaxConcurrentProcessing()),
 		UpstreamFetchTimeout: loader.durationVal(
 			keyUpstreamFetchTimeout, DefaultUpstreamFetchTimeout),
 		UpstreamMaxResponseSize: loader.int64Val(

internal/imageprocessor/imageprocessor.go

@@ -7,11 +7,11 @@ import (
 	"errors"
 	"fmt"
 	"io"
-	"runtime"
 	"sync"
 	"time"
 
 	"github.com/davidbyttow/govips/v2/vips"
+	"sneak.berlin/go/pixa/internal/config"
 )
 
 // vipsOnce ensures vips is initialized exactly once.
@@ -20,11 +20,10 @@ import (
 var vipsOnce sync.Once
 
 // initVips initializes libvips with quiet logging, one worker thread per
-// image and no operation cache. Process already works on one image per CPU
+// image and no operation cache. Each request decodes different source
-// by default, so more threads per image would only compete for the CPUs.
+// bytes, so the operation cache would rarely be hit and would hold memory
-// Each request decodes different source bytes, so the operation cache
+// outside MaxConcurrentProcessing; repeated requests are served from
-// would rarely be hit and would hold memory outside MaxConcurrentProcessing;
+// pixa's disk cache instead.
-// repeated requests are served from pixa's disk cache instead.
 func initVips() {
 	vipsOnce.Do(func() {
 		vips.LoggingSettings(nil, vips.LogLevelError)
@@ -179,7 +178,7 @@ type Params struct {
 	// If <= 0, DefaultMaxInputBytes is used.
 	MaxInputBytes int64
 	// MaxConcurrentProcessing is the most images processed at once.
-	// If <= 0, the number of CPUs Go uses (runtime.GOMAXPROCS(0)) is used.
+	// If <= 0, config.DefaultMaxConcurrentProcessing() is used.
 	MaxConcurrentProcessing int
 }
 
@@ -195,7 +194,7 @@ func New(params Params) *ImageProcessor {
 
 	maxConcurrentProcessing := params.MaxConcurrentProcessing
 	if maxConcurrentProcessing <= 0 {
-		maxConcurrentProcessing = runtime.GOMAXPROCS(0)
+		maxConcurrentProcessing = config.DefaultMaxConcurrentProcessing()
 	}
 
 	return &ImageProcessor{

internal/imageprocessor/max_concurrent_processing_internal_test.go

@@ -115,14 +115,14 @@ func waitForEntries(t *testing.T, entered <-chan struct{}, count int) {
 	}
 }
 
-func TestNewDefaultsMaxConcurrentProcessingToCPUs(t *testing.T) {
+func TestNewDefaultsMaxConcurrentProcessingToHalfTheCPUs(t *testing.T) {
 	t.Parallel()
 
 	for _, limit := range []int{0, -1} {
 		proc := New(Params{MaxConcurrentProcessing: limit})
-		if got := cap(proc.processingSemaphore); got != runtime.GOMAXPROCS(0) {
+		if got := cap(proc.processingSemaphore); got != max(runtime.GOMAXPROCS(0)/2, 1) {
-			t.Errorf("MaxConcurrentProcessing %d: %d slots, want %d, one per CPU",
+			t.Errorf("MaxConcurrentProcessing %d: %d slots, want %d, half the CPUs",
-				limit, got, runtime.GOMAXPROCS(0))
+				limit, got, max(runtime.GOMAXPROCS(0)/2, 1))
 		}
 	}
 
@@ -132,6 +132,30 @@ func TestNewDefaultsMaxConcurrentProcessingToCPUs(t *testing.T) {
 	}
 }
 
+// TestNewDefaultIsHalfTheCPUsAndAtLeastOne checks that New, given no
+// MaxConcurrentProcessing, makes one slot for every two CPUs Go uses,
+// rounded down, and one slot when Go uses one CPU.
+//
+//nolint:paralleltest // it sets GOMAXPROCS, which other tests read
+func TestNewDefaultIsHalfTheCPUsAndAtLeastOne(t *testing.T) {
+	cpusBefore := runtime.GOMAXPROCS(0)
+
+	t.Cleanup(func() { runtime.GOMAXPROCS(cpusBefore) })
+
+	for _, tc := range []struct{ cpus, want int }{
+		{1, 1}, {2, 1}, {3, 1}, {8, 4},
+	} {
+		runtime.GOMAXPROCS(tc.cpus)
+
+		got := cap(New(Params{}).processingSemaphore)
+		t.Logf("GOMAXPROCS %d: %d slots", tc.cpus, got)
+
+		if got != tc.want {
+			t.Errorf("GOMAXPROCS %d: %d slots, want %d", tc.cpus, got, tc.want)
+		}
+	}
+}
+
 // TestProcessNeverExceedsMaxConcurrentProcessing starts more Process calls
 // than MaxConcurrentProcessing allows and holds the first ones inside
 // Process until the test lets them go. No more than the limit may be

internal/imgcache/cache.go

@@ -660,10 +660,10 @@ func (c *Cache) readHitAndMissCounts(ctx context.Context, stats *CacheStats) err
 // hold: it reads the content type from the variant's .meta file and
 // keeps it in metaCache, unless a StoreVariant has put one there
 // meanwhile, as the store's is newer. A read that finds no .meta file,
-// as one can between a store's writing of the variant file and of its
+// as one can when the store's .meta write failed or an eviction removed
-// .meta file, serves application/octet-stream and keeps nothing, so
+// the file after the read opened the variant, serves
-// metaCache only ever holds a type read from a .meta file or passed to
+// application/octet-stream and keeps nothing, so metaCache only ever
-// StoreVariant.
+// holds a type read from a .meta file or passed to StoreVariant.
 func (c *Cache) loadVariantWithMeta(
 	cacheKey VariantKey,
 ) (io.ReadCloser, int64, string, error) {

internal/imgcache/eviction.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"database/sql"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io/fs"
 	"os"
@@ -560,10 +561,11 @@ func (c *Cache) runCleanExpired(ctx context.Context) {
 // (or whose accounting insert failed, e.g. StoreVariant's best-effort
 // insert under transient DB contention), drops accounting rows whose
 // files are missing, removes source blob files the database does not
-// know (and rows whose files are gone), sweeps stale temp files left
+// know (and rows whose files are gone), removes source and variant files
-// behind by crashed writes, and last checks the total cache usage
+// that are empty or not the size their rows record, sweeps stale temp
-// against the tables. It reads the tables a page at a time. Running it
+// files left behind by crashed writes, and last checks the total cache
-// periodically, not just once, bounds how long such drift can
+// usage against the tables. It reads the tables a page at a time.
+// Running it periodically, not just once, bounds how long such drift can
 // accumulate unaccounted for on a long-running process to one eviction
 // interval. Once ctx is cancelled, it stops at the next file or row and
 // returns ctx's error.
@@ -595,8 +597,9 @@ func (c *Cache) reconcileAccounting(ctx context.Context) error {
 	return c.reconcileUsageTotal(ctx)
 }
 
-// reconcileVariantFiles walks the variant storage directory, adopting
+// reconcileVariantFiles walks the variant storage directory, checking
-// files without accounting rows and sweeping stale temp files.
+// each variant file against its accounting row (reconcileVariantFile)
+// and sweeping stale temp files.
 func (c *Cache) reconcileVariantFiles(ctx context.Context) error {
 	return filepath.WalkDir(
 		c.variants.baseDir,
@@ -621,38 +624,55 @@ func (c *Cache) reconcileVariantFiles(ctx context.Context) error {
 				return nil
 			}
 
-			return c.adoptVariantFile(ctx, path, entry, VariantKey(name))
+			return c.reconcileVariantFile(ctx, path, entry, VariantKey(name))
 		},
 	)
 }
 
-// adoptVariantFile inserts an accounting row for a variant file that
+// reconcileVariantFile checks one variant file against its accounting
-// has none, using the file's size and modification time.
+// row. A file without a row is adopted into the accounting, unless it is
-func (c *Cache) adoptVariantFile(
+// empty. A file that is empty, or not the size its row records, as a
+// crash or a restored snapshot can leave it, is removed with its row and
+// a warning naming it, so the variant is converted again.
+func (c *Cache) reconcileVariantFile(
 	ctx context.Context, path string, entry fs.DirEntry, cacheKey VariantKey,
 ) error {
-	var rowExists int
-
-	err := c.db.QueryRowContext(ctx,
-		`SELECT COUNT(*) FROM variant_content WHERE cache_key = ?`, string(cacheKey),
-	).Scan(&rowExists)
-	if err != nil {
-		return fmt.Errorf("failed to check variant accounting row: %w", err)
-	}
-
-	if rowExists > 0 {
-		return nil
-	}
-
 	info, err := entry.Info()
 	if err != nil {
 		return fmt.Errorf("failed to stat variant file: %w", err)
 	}
 
+	var recordedSize int64
+
+	err = c.db.QueryRowContext(ctx,
+		`SELECT size_bytes FROM variant_content WHERE cache_key = ?`, string(cacheKey),
+	).Scan(&recordedSize)
+
+	switch {
+	case errors.Is(err, sql.ErrNoRows) && info.Size() > 0:
+		return c.adoptVariantFile(ctx, path, info, cacheKey)
+	case err != nil && !errors.Is(err, sql.ErrNoRows):
+		return fmt.Errorf("failed to check variant accounting row: %w", err)
+	case info.Size() > 0 && info.Size() == recordedSize:
+		return nil
+	}
+
+	// The file is empty, or not the size its row records
+	c.log.Warn("removing damaged cached variant file", "path", path,
+		"file_bytes", info.Size(), "recorded_bytes", recordedSize)
+
+	return c.evictVariant(ctx, cacheKey)
+}
+
+// adoptVariantFile inserts an accounting row for a variant file that
+// has none, using the file's size and modification time.
+func (c *Cache) adoptVariantFile(
+	ctx context.Context, path string, info fs.FileInfo, cacheKey VariantKey,
+) error {
 	modTime := info.ModTime().UTC().Format(sqliteTimestampLayout)
 	contentType := c.variantContentTypeFromSidecar(path)
 
-	_, err = c.db.ExecContext(ctx, `
+	_, err := c.db.ExecContext(ctx, `
 		INSERT INTO variant_content
 			(cache_key, size_bytes, content_type, created_at, last_accessed_at)
 		VALUES (?, ?, ?, ?, ?)
@@ -771,10 +791,9 @@ func queryStringColumn[T ~string](
 	return values, nil
 }
 
-// reconcileSourceFiles walks the source content directory, removing
+// reconcileSourceFiles walks the source content directory, checking
-// blob files the database does not track (they are unreachable: source
+// each blob file against its source_content row (reconcileSourceFile)
-// lookups always go through source_metadata) and sweeping stale temp
+// and sweeping stale temp files.
-// files.
 func (c *Cache) reconcileSourceFiles(ctx context.Context) error {
 	return filepath.WalkDir(
 		c.srcContent.baseDir,
@@ -795,31 +814,55 @@ func (c *Cache) reconcileSourceFiles(ctx context.Context) error {
 				return nil
 			}
 
-			return c.removeUntrackedSourceFile(ctx, path, ContentHash(name))
+			return c.reconcileSourceFile(ctx, path, entry, ContentHash(name))
 		},
 	)
 }
 
+// reconcileSourceFile checks one source blob file against its
+// source_content row. A file without a row is removed: it is
+// unreachable, as source lookups always go through source_metadata. A
+// file that is empty, or not the size its row records, as a crash or a
+// restored snapshot can leave it, is removed with its rows and a warning
+// naming it, so the source is fetched again.
+func (c *Cache) reconcileSourceFile(
+	ctx context.Context, path string, entry fs.DirEntry, contentHash ContentHash,
+) error {
+	info, err := entry.Info()
+	if err != nil {
+		return fmt.Errorf("failed to stat source content file: %w", err)
+	}
+
+	var recordedSize int64
+
+	err = c.db.QueryRowContext(ctx,
+		`SELECT size_bytes FROM source_content WHERE content_hash = ?`,
+		string(contentHash),
+	).Scan(&recordedSize)
+
+	switch {
+	case errors.Is(err, sql.ErrNoRows):
+		return c.removeUntrackedSourceFile(ctx, path, contentHash)
+	case err != nil:
+		return fmt.Errorf("failed to check source content row: %w", err)
+	case info.Size() > 0 && info.Size() == recordedSize:
+		return nil
+	}
+
+	// The file is empty, or not the size its row records
+	c.log.Warn("removing damaged cached source file", "path", path,
+		"file_bytes", info.Size(), "recorded_bytes", recordedSize)
+
+	return c.evictSourceBlob(ctx, contentHash)
+}
+
 // removeUntrackedSourceFile deletes a source blob file that has no
 // source_content row. Any source_metadata rows referencing the hash
 // are removed first so no row ever points at a deleted file.
 func (c *Cache) removeUntrackedSourceFile(
 	ctx context.Context, path string, contentHash ContentHash,
 ) error {
-	var rowExists int
+	_, err := c.db.ExecContext(ctx,
-
-	err := c.db.QueryRowContext(ctx,
-		`SELECT COUNT(*) FROM source_content WHERE content_hash = ?`, string(contentHash),
-	).Scan(&rowExists)
-	if err != nil {
-		return fmt.Errorf("failed to check source content row: %w", err)
-	}
-
-	if rowExists > 0 {
-		return nil
-	}
-
-	_, err = c.db.ExecContext(ctx,
 		`DELETE FROM source_metadata WHERE content_hash = ?`, string(contentHash))
 	if err != nil {
 		return fmt.Errorf("failed to delete metadata rows for untracked blob: %w", err)

internal/imgcache/partial_cache_files_internal_test.go

@@ -0,0 +1,359 @@
+package imgcache
+
+import (
+	"bytes"
+	"crypto/sha256"
+	"encoding/hex"
+	"io"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// getImage calls Get for req and returns the image it answered with and
+// its cache status.
+func getImage(t *testing.T, svc *Service, req *ImageRequest) ([]byte, CacheStatus) {
+	t.Helper()
+
+	resp, err := svc.Get(t.Context(), req)
+	if err != nil {
+		t.Fatalf("Get() of quality %d error = %v", req.Quality, err)
+	}
+
+	defer func() { _ = resp.Content.Close() }()
+
+	body, err := io.ReadAll(resp.Content)
+	if err != nil {
+		t.Fatalf("reading the image of quality %d: %v", req.Quality, err)
+	}
+
+	return body, resp.CacheStatus
+}
+
+// fileSize returns the size of the file at path.
+func fileSize(t *testing.T, path string) int64 {
+	t.Helper()
+
+	info, err := os.Stat(path)
+	if err != nil {
+		t.Fatalf("stat %s: %v", path, err)
+	}
+
+	return info.Size()
+}
+
+// truncateFile cuts the file at path down to size bytes.
+func truncateFile(t *testing.T, path string, size int64) {
+	t.Helper()
+
+	t.Logf("cutting %s from %d to %d bytes",
+		filepath.Base(path), fileSize(t, path), size)
+
+	err := os.Truncate(path, size)
+	if err != nil {
+		t.Fatalf("truncating %s: %v", path, err)
+	}
+}
+
+// cachedSourcePath returns the path of the cached source file req uses.
+func cachedSourcePath(t *testing.T, svc *Service, req *ImageRequest) string {
+	t.Helper()
+
+	contentHash, _, err := svc.cache.LookupSource(t.Context(), req)
+	if err != nil || contentHash == "" {
+		t.Fatalf("LookupSource() = %q, %v; want the cached source", contentHash, err)
+	}
+
+	return svc.cache.srcContent.hashToPath(contentHash)
+}
+
+// reconcileAsAtStartup runs the cache's reconciliation pass, as pixa does
+// when it starts, and returns what the pass logged.
+func reconcileAsAtStartup(t *testing.T, cache *Cache) string {
+	t.Helper()
+
+	var logBuf bytes.Buffer
+
+	cache.log = slog.New(slog.NewJSONHandler(&logBuf, nil))
+
+	err := cache.reconcileAccounting(t.Context())
+	if err != nil {
+		t.Fatalf("reconcileAccounting() error = %v", err)
+	}
+
+	t.Logf("reconciliation log: %s", logBuf.String())
+
+	return logBuf.String()
+}
+
+// assertWarningNames checks that logs hold a warning naming the file at
+// path.
+func assertWarningNames(t *testing.T, logs, path string) {
+	t.Helper()
+
+	for line := range strings.Lines(logs) {
+		if strings.Contains(line, `"level":"WARN"`) &&
+			strings.Contains(line, `"path":"`+path+`"`) {
+			return
+		}
+	}
+
+	t.Errorf("reconciliation logged no warning naming %s", path)
+}
+
+// TestService_Get_FetchesAgainForEmptyCachedSource empties the cached source
+// file, as a crash or a restored snapshot could, and runs the reconciliation
+// pass, which must remove it with a warning naming it. It then asks for two
+// new sizes of the source: the first must fetch it again and cache it whole,
+// so that the second is made from the cache without a third fetch.
+func TestService_Get_FetchesAgainForEmptyCachedSource(t *testing.T) {
+	t.Parallel()
+
+	svc, fixtures := SetupTestService(t)
+
+	getImage(t, svc, photoVariant(fixtures, 85, FitCover))
+
+	path := cachedSourcePath(t, svc, photoVariant(fixtures, 85, FitCover))
+	wholeSize := fileSize(t, path)
+
+	truncateFile(t, path, 0)
+	assertWarningNames(t, reconcileAsAtStartup(t, svc.cache), path)
+
+	getImage(t, svc, photoVariant(fixtures, 40, FitCover))
+	getImage(t, svc, photoVariant(fixtures, 30, FitCover))
+
+	fetches := readCacheStatsCounters(t, svc.cache).upstreamFetchCount
+	t.Logf("upstream fetches = %d", fetches)
+
+	if fetches != 2 {
+		t.Errorf("upstream fetches = %d, want 2: the empty cached source "+
+			"was kept and fetched again on every new size", fetches)
+	}
+
+	size := fileSize(t, path)
+	if size != wholeSize {
+		t.Errorf("cached source is %d bytes, want %d", size, wholeSize)
+	}
+}
+
+// TestService_Get_FetchesAgainForPartialCachedSource cuts the cached source
+// file to half its size, as a crash or a restored snapshot could, and runs
+// the reconciliation pass, which must remove it with a warning naming it.
+// It then asks for two new sizes of the source: the first must fetch it
+// again rather than fail to decode the partial file, and the second is
+// made from the cache.
+func TestService_Get_FetchesAgainForPartialCachedSource(t *testing.T) {
+	t.Parallel()
+
+	svc, fixtures := SetupTestService(t)
+
+	getImage(t, svc, photoVariant(fixtures, 85, FitCover))
+
+	path := cachedSourcePath(t, svc, photoVariant(fixtures, 85, FitCover))
+	wholeSize := fileSize(t, path)
+
+	truncateFile(t, path, wholeSize/2)
+	assertWarningNames(t, reconcileAsAtStartup(t, svc.cache), path)
+
+	getImage(t, svc, photoVariant(fixtures, 40, FitCover))
+	getImage(t, svc, photoVariant(fixtures, 30, FitCover))
+
+	fetches := readCacheStatsCounters(t, svc.cache).upstreamFetchCount
+	t.Logf("upstream fetches = %d", fetches)
+
+	if fetches != 2 {
+		t.Errorf("upstream fetches = %d, want 2", fetches)
+	}
+
+	size := fileSize(t, path)
+	if size != wholeSize {
+		t.Errorf("cached source is %d bytes, want %d", size, wholeSize)
+	}
+}
+
+// assertConvertedAgain caches a variant of the photo, cuts its file to size
+// bytes and runs the reconciliation pass, which must remove it with a
+// warning naming it. It then checks that Get converts the image again
+// instead of serving the file, and that the next Get serves the whole
+// variant from the cache.
+func assertConvertedAgain(t *testing.T, size int64) {
+	t.Helper()
+
+	svc, fixtures := SetupTestService(t)
+	req := photoVariant(fixtures, 85, FitCover)
+
+	want, _ := getImage(t, svc, req)
+
+	path := svc.cache.variants.keyToPath(CacheKey(req))
+
+	truncateFile(t, path, size)
+	assertWarningNames(t, reconcileAsAtStartup(t, svc.cache), path)
+
+	got, status := getImage(t, svc, req)
+	t.Logf("after the cut: %d bytes, %s", len(got), status)
+
+	if status != CacheMiss || !bytes.Equal(got, want) {
+		t.Errorf("after the cut: Get() = %d bytes, %s; want the %d bytes "+
+			"converted again, %s", len(got), status, len(want), CacheMiss)
+	}
+
+	got, status = getImage(t, svc, req)
+	t.Logf("next request: %d bytes, %s", len(got), status)
+
+	if status != CacheHit || !bytes.Equal(got, want) {
+		t.Errorf("next request: Get() = %d bytes, %s; want the %d bytes, %s",
+			len(got), status, len(want), CacheHit)
+	}
+}
+
+// TestService_Get_ConvertsAgainForEmptyCachedVariant empties a cached
+// variant file, as a crash or a restored snapshot could, and checks that it
+// is converted again rather than served empty.
+func TestService_Get_ConvertsAgainForEmptyCachedVariant(t *testing.T) {
+	t.Parallel()
+
+	assertConvertedAgain(t, 0)
+}
+
+// TestService_Get_ConvertsAgainForPartialCachedVariant cuts a cached
+// variant file to ten bytes, as a crash or a restored snapshot could, and
+// checks that it is converted again rather than served cut short.
+func TestService_Get_ConvertsAgainForPartialCachedVariant(t *testing.T) {
+	t.Parallel()
+
+	assertConvertedAgain(t, 10)
+}
+
+// TestReconciliationRemovesEmptyVariantFileWithoutRow checks that the
+// reconciliation pass removes an empty variant file that has no
+// accounting row, with a warning naming it, instead of adding a row for
+// it that would have it served empty.
+func TestReconciliationRemovesEmptyVariantFileWithoutRow(t *testing.T) {
+	t.Parallel()
+
+	cache, _ := newEvictionTestCache(t, 1<<30)
+
+	_, err := cache.variants.Store(testVariantKeyOne, bytes.NewReader(nil),
+		testContentTypeWebP)
+	if err != nil {
+		t.Fatalf("failed to store an empty variant file: %v", err)
+	}
+
+	logs := reconcileAsAtStartup(t, cache)
+
+	assertWarningNames(t, logs, cache.variants.keyToPath(testVariantKeyOne))
+
+	if cache.variants.Exists(testVariantKeyOne) {
+		t.Error("the empty variant file is still on disk")
+	}
+
+	rows := countRows(t, cache,
+		`SELECT COUNT(*) FROM variant_content WHERE cache_key = ?`,
+		string(testVariantKeyOne))
+	if rows != 0 {
+		t.Errorf("the empty variant file got %d accounting rows, want 0", rows)
+	}
+}
+
+// TestVariantStorage_StoreReplacesMetaFile stores a variant over one
+// already stored and checks that Store puts a new .meta file in place of
+// the old one, as it does the image, rather than writing into the old
+// file, which a crash could leave empty. The old .meta file is given a
+// second name before the store, so that it can be told apart from the new
+// one after it.
+func TestVariantStorage_StoreReplacesMetaFile(t *testing.T) {
+	t.Parallel()
+
+	storage, key := storeTestVariant(t)
+	metaPath := storage.keyToPath(key) + variantMetaSuffix
+	oldMetaPath := metaPath + ".old"
+
+	err := os.Link(metaPath, oldMetaPath)
+	if err != nil {
+		t.Fatalf("giving the .meta file a second name: %v", err)
+	}
+
+	_, err = storage.Store(key, bytes.NewReader([]byte("new variant data")),
+		testContentTypeJPEG)
+	if err != nil {
+		t.Fatalf("Store() error = %v", err)
+	}
+
+	oldMeta, err := os.Stat(oldMetaPath)
+	if err != nil {
+		t.Fatalf("stat of the old .meta file: %v", err)
+	}
+
+	newMeta, err := os.Stat(metaPath)
+	if err != nil {
+		t.Fatalf("stat of the new .meta file: %v", err)
+	}
+
+	if os.SameFile(oldMeta, newMeta) {
+		t.Error("Store() wrote into the old .meta file instead of putting " +
+			"a new one in its place")
+	}
+
+	reader, _, contentType, err := storage.LoadWithMeta(key)
+	if err != nil {
+		t.Fatalf("LoadWithMeta() error = %v", err)
+	}
+
+	_ = reader.Close()
+
+	t.Logf("content type after the store: %s", contentType)
+
+	if contentType != testContentTypeJPEG {
+		t.Errorf("content type = %q, want %q", contentType, testContentTypeJPEG)
+	}
+}
+
+// TestContentStorage_StoreHashedReplacesFileOfAnotherSize checks that a
+// store over a file of another size at the content's path, such as a
+// partial file a crash left, replaces it with the content.
+func TestContentStorage_StoreHashedReplacesFileOfAnotherSize(t *testing.T) {
+	t.Parallel()
+
+	storage, err := NewContentStorage(t.TempDir())
+	if err != nil {
+		t.Fatalf("NewContentStorage() error = %v", err)
+	}
+
+	content := []byte("the whole source image")
+	sum := sha256.Sum256(content)
+	hash := ContentHash(hex.EncodeToString(sum[:]))
+	path := storage.hashToPath(hash)
+
+	err = os.MkdirAll(filepath.Dir(path), StorageDirPerm)
+	if err != nil {
+		t.Fatalf("creating the content's directory: %v", err)
+	}
+
+	err = os.WriteFile(path, content[:5], StorageFilePerm)
+	if err != nil {
+		t.Fatalf("writing a partial file: %v", err)
+	}
+
+	_, err = storage.StoreHashed(hash, content)
+	if err != nil {
+		t.Fatalf("StoreHashed() error = %v", err)
+	}
+
+	stored, err := storage.Load(hash)
+	if err != nil {
+		t.Fatalf("Load() error = %v", err)
+	}
+
+	defer func() { _ = stored.Close() }()
+
+	got, err := io.ReadAll(stored)
+	if err != nil {
+		t.Fatalf("reading the stored content: %v", err)
+	}
+
+	if !bytes.Equal(got, content) {
+		t.Errorf("stored file = %q, want %q", got, content)
+	}
+}

internal/imgcache/service.go

@@ -53,7 +53,7 @@ type ServiceConfig struct {
 	// Allowlist is the list of hosts that don't require signatures
 	Allowlist []string
 	// MaxConcurrentProcessing is the most images processed at once; zero
-	// uses the image processor's default, one per CPU
+	// uses the image processor's default, half the CPUs
 	MaxConcurrentProcessing int
 	// Logger for logging
 	Logger *slog.Logger

internal/imgcache/storage.go

@@ -80,8 +80,8 @@ func (s *ContentStorage) Store(r io.Reader) (ContentHash, int64, error) {
 // from hash, without recomputing it. Callers that already know the
 // hash before writing (e.g. because they must hold a hash-keyed lock
 // across the whole store operation) use this instead of Store. Like
-// Store, it is idempotent: content already on disk at that path is
+// Store, it is idempotent: a file of the content's size already at that
-// left untouched.
+// path is left untouched, and a file of another size is replaced.
 func (s *ContentStorage) StoreHashed(
 	hash ContentHash, data []byte,
 ) (int64, error) {
@@ -152,28 +152,35 @@ func (s *ContentStorage) Exists(hash ContentHash) bool {
 	return err == nil
 }
 
-// writeIfAbsent writes data to the path derived from hash, unless
+// writeIfAbsent writes data with writeFileSynced to the path derived
-// content already exists there, via a temp-file-plus-rename so
+// from hash, unless a file of the data's size is already there.
-// concurrent readers never observe a partial file.
 func (s *ContentStorage) writeIfAbsent(hash ContentHash, data []byte) error {
 	// Build path: <basedir>/<ab>/<cd>/<hash>
 	path := s.hashToPath(hash)
 
-	// Check if already exists
+	// A file of another size there is damaged, and is replaced
-	_, statErr := os.Stat(path)
+	info, statErr := os.Stat(path)
-	if statErr == nil {
+	if statErr == nil && info.Size() == int64(len(data)) {
 		return nil
 	}
 
-	// Create directory structure
+	err := os.MkdirAll(filepath.Dir(path), StorageDirPerm)
-	dir := filepath.Dir(path)
-
-	err := os.MkdirAll(dir, StorageDirPerm)
 	if err != nil {
 		return fmt.Errorf("failed to create directory: %w", err)
 	}
 
-	// Write to temp file first, then rename for atomicity
+	return writeFileSynced(path, data)
+}
+
+// writeFileSynced writes data to path through a temp file in the same
+// directory: the temp file is synced to disk, renamed to path, and the
+// directory synced after, so that neither a concurrent reader nor a
+// crash finds a partial file under path. When the directory sync fails,
+// the file is removed from path, so that a failed write leaves no file
+// there.
+func writeFileSynced(path string, data []byte) error {
+	dir := filepath.Dir(path)
+
 	tmpFile, err := os.CreateTemp(dir, ".tmp-*")
 	if err != nil {
 		return fmt.Errorf("failed to create temp file: %w", err)
@@ -181,18 +188,20 @@ func (s *ContentStorage) writeIfAbsent(hash ContentHash, data []byte) error {
 
 	tmpPath := tmpFile.Name()
 
-	// Each failure path below unlinks the temp file explicitly. This
-	// replaces a deferred cleanup that read a named result, which the
-	// canonical config does not permit; the set of paths that remove
-	// tmpPath, and the order relative to Close, is unchanged. This
-	// mirrors how MetadataStorage.Store and VariantStorage.Store below
-	// already express the same cleanup.
 	_, err = tmpFile.Write(data)
 	if err != nil {
 		_ = tmpFile.Close()
 		_ = os.Remove(tmpPath)
 
-		return fmt.Errorf("failed to write content: %w", err)
+		return fmt.Errorf("failed to write temp file: %w", err)
+	}
+
+	err = tmpFile.Sync()
+	if err != nil {
+		_ = tmpFile.Close()
+		_ = os.Remove(tmpPath)
+
+		return fmt.Errorf("failed to sync temp file: %w", err)
 	}
 
 	err = tmpFile.Close()
@@ -202,7 +211,6 @@ func (s *ContentStorage) writeIfAbsent(hash ContentHash, data []byte) error {
 		return fmt.Errorf("failed to close temp file: %w", err)
 	}
 
-	// Atomic rename
 	err = os.Rename(filepath.Clean(tmpPath), filepath.Clean(path))
 	if err != nil {
 		_ = os.Remove(tmpPath)
@@ -210,6 +218,31 @@ func (s *ContentStorage) writeIfAbsent(hash ContentHash, data []byte) error {
 		return fmt.Errorf("failed to rename temp file: %w", err)
 	}
 
+	err = syncDirectory(dir)
+	if err != nil {
+		_ = os.Remove(path)
+
+		return err
+	}
+
+	return nil
+}
+
+// syncDirectory syncs the directory dir to disk, so that a file just
+// renamed into it is still there after a crash.
+func syncDirectory(dir string) error {
+	d, err := os.Open(filepath.Clean(dir))
+	if err != nil {
+		return fmt.Errorf("failed to open directory: %w", err)
+	}
+
+	err = d.Sync()
+	_ = d.Close()
+
+	if err != nil {
+		return fmt.Errorf("failed to sync directory: %w", err)
+	}
+
 	return nil
 }
 
@@ -427,7 +460,9 @@ func NewVariantStorage(baseDir string, log *slog.Logger) (*VariantStorage, error
 	return &VariantStorage{baseDir: baseDir, log: log}, nil
 }
 
-// Store writes content and metadata to storage at the given key.
+// Store writes content and metadata to storage at the given key, each
+// with writeFileSynced. The .meta file is written first, so that a crash
+// cannot leave the content in place without it.
 func (s *VariantStorage) Store(
 	key VariantKey, r io.Reader, contentType string,
 ) (int64, error) {
@@ -440,46 +475,11 @@ func (s *VariantStorage) Store(
 	path := s.keyToPath(key)
 	metaPath := path + ".meta"
 
-	// Create directory structure
+	err = os.MkdirAll(filepath.Dir(path), StorageDirPerm)
-	dir := filepath.Dir(path)
-
-	err = os.MkdirAll(dir, StorageDirPerm)
 	if err != nil {
 		return 0, fmt.Errorf("failed to create directory: %w", err)
 	}
 
-	// Write content to temp file first, then rename for atomicity
-	tmpFile, err := os.CreateTemp(dir, ".tmp-*")
-	if err != nil {
-		return 0, fmt.Errorf("failed to create temp file: %w", err)
-	}
-
-	tmpPath := tmpFile.Name()
-
-	_, err = tmpFile.Write(data)
-	if err != nil {
-		_ = tmpFile.Close()
-		_ = os.Remove(tmpPath)
-
-		return 0, fmt.Errorf("failed to write content: %w", err)
-	}
-
-	err = tmpFile.Close()
-	if err != nil {
-		_ = os.Remove(tmpPath)
-
-		return 0, fmt.Errorf("failed to close temp file: %w", err)
-	}
-
-	// Atomic rename content
-	err = os.Rename(filepath.Clean(tmpPath), filepath.Clean(path))
-	if err != nil {
-		_ = os.Remove(tmpPath)
-
-		return 0, fmt.Errorf("failed to rename temp file: %w", err)
-	}
-
-	// Write metadata
 	meta := VariantMeta{
 		ContentType: contentType,
 		Size:        size,
@@ -491,13 +491,19 @@ func (s *VariantStorage) Store(
 		return 0, fmt.Errorf("failed to marshal metadata: %w", err)
 	}
 
-	// Metadata write failure is non-fatal; content is already stored.
+	// Metadata write failure is non-fatal: the content is then served as
-	err = os.WriteFile(metaPath, metaData, StorageFilePerm)
+	// application/octet-stream.
+	err = writeFileSynced(metaPath, metaData)
 	if err != nil {
 		s.log.Warn("failed to write variant metadata sidecar",
 			"path", metaPath, "error", err)
 	}
 
+	err = writeFileSynced(path, data)
+	if err != nil {
+		return 0, err
+	}
+
 	return size, nil
 }