feat(cache): add shared S3-compatible storage (#1244)

Store the cache used by `actions/cache` in an S3-compatible bucket such as AWS S3 or RustFS, with no `cache-server`. Runners with the same settings use the same cache.

A cache saved on one runner can be restored on the others within 30 seconds. The runner talks to the bucket over the S3 REST API with SigV4, without an SDK. Lifecycle rules on the bucket expire old caches.

*Description rewritten by Claude.*

---------

Co-authored-by: silverwind <[email protected]>
Reviewed-on: https://gitea.com/gitea/runner/pulls/1244
Reviewed-by: silverwind <[email protected]>
Co-authored-by: bircni <[email protected]>

act/artifactcache/handler.go

@@ -10,6 +10,7 @@ import (
 	"crypto/hmac"
 	"crypto/rand"
 	"crypto/sha256"
+	"encoding/binary"
 	"encoding/hex"
 	"encoding/json/v2"
 	"errors"
@@ -78,7 +79,8 @@ type credEntry struct {
 
 type Handler struct {
 	dir      string
-	storage  *Storage
+	storage  cacheStorage
+	s3       *s3Storage
 	router   *httprouter.Router
 	listener net.Listener
 	port     int
@@ -88,6 +90,9 @@ type Handler struct {
 	gcing atomic.Bool
 	gcAt  time.Time
 
+	stopSync context.CancelFunc
+	syncing  sync.WaitGroup
+
 	outboundIP string
 
 	// internalSecret guards /_internal/{register,revoke}. When set, a remote
@@ -117,6 +122,7 @@ type Options struct {
 	Dir        string
 	OutboundIP string
 	Port       uint16
+	S3         *S3Options
 	Upstream   string
 
 	// InternalSecret, when non-empty, enables a control-plane API at
@@ -185,6 +191,12 @@ func StartHandler(opts Options) (*Handler, error) {
 		return nil, err
 	}
 	h.storage = storage
+	if opts.S3 != nil {
+		if h.s3, err = newS3Storage(storage, *opts.S3); err != nil {
+			return nil, err
+		}
+		h.storage = h.s3
+	}
 
 	secret, err := loadOrCreateSecret(dir)
 	if err != nil {
@@ -216,6 +228,11 @@ func StartHandler(opts Options) (*Handler, error) {
 	if err := h.serve(opts.Port, router); err != nil {
 		return nil, err
 	}
+	if h.s3 != nil {
+		ctx, cancel := context.WithCancel(context.Background())
+		h.stopSync = cancel
+		h.syncing.Go(func() { h.syncMetadata(ctx) })
+	}
 	return h, nil
 }
 
@@ -380,6 +397,10 @@ func (h *Handler) Close() error {
 	if h == nil {
 		return nil
 	}
+	if h.stopSync != nil {
+		h.stopSync()
+		h.syncing.Wait()
+	}
 	var retErr error
 	if h.server != nil {
 		err := h.server.Close()
@@ -413,6 +434,72 @@ func (h *Handler) openDB() (*bolthold.Store, error) {
 	})
 }
 
+// syncMetadata mirrors the bucket into the store until ctx ends, so no request waits on S3 for it.
+func (h *Handler) syncMetadata(ctx context.Context) {
+	ticker := time.NewTicker(metadataSyncInterval)
+	defer ticker.Stop()
+	for {
+		if err := h.importMetadata(ctx); err != nil && ctx.Err() == nil {
+			h.logger.Warnf("sync S3 cache: %v", err)
+		}
+		select {
+		case <-ctx.Done():
+			return
+		case <-ticker.C:
+		}
+	}
+}
+
+// importMetadata lists after snapshotting the store, so an entry committed meanwhile, published before it is stored, stays.
+func (h *Handler) importMetadata(ctx context.Context) error {
+	db, err := h.openDB()
+	if err != nil {
+		return err
+	}
+	local := h.caches(db, bolthold.Where("Complete").Eq(true).Index("Complete"))
+	db.Close()
+
+	published, err := h.s3.published(ctx)
+	if err != nil {
+		return err
+	}
+	var gone []*Cache
+	for _, cache := range local {
+		if !published[cache.ID] {
+			gone = append(gone, cache)
+		}
+		delete(published, cache.ID)
+	}
+	added := make([]*Cache, 0, len(published))
+	for id := range published {
+		cache, err := h.s3.fetch(ctx, id)
+		if err != nil {
+			return err
+		}
+		added = append(added, cache)
+	}
+
+	db, err = h.openDB()
+	if err != nil {
+		return err
+	}
+	defer db.Close()
+	return db.Bolt().Update(func(tx *bbolt.Tx) error {
+		for _, cache := range added {
+			if err := db.TxUpsert(tx, cache.ID, cache); err != nil {
+				return err
+			}
+		}
+		for _, cache := range gone {
+			_ = h.s3.Storage.Remove(cache.ID) // a blob kept on disk before the runner used S3
+			if err := db.TxDelete(tx, cache.ID, cache); err != nil && !errors.Is(err, bolthold.ErrNotFound) {
+				return err
+			}
+		}
+		return nil
+	})
+}
+
 // GET /_apis/artifactcache/cache
 func (h *Handler) find(w http.ResponseWriter, r *http.Request, _ httprouter.Params) {
 	cred := credFromContext(r.Context())
@@ -455,7 +542,7 @@ func (h *Handler) lookupCache(db *bolthold.Store, repo string, keys []string, ve
 		return nil, err
 	}
 	if !ok {
-		_ = db.Delete(cache.ID, cache)
+		h.deleteCache(db, cache)
 		return nil, nil //nolint:nilnil // absence is not an error here
 	}
 	// Handing out a download URL counts as access, or eviction could drop the entry between
@@ -487,7 +574,7 @@ func (h *Handler) reserve(w http.ResponseWriter, r *http.Request, _ httprouter.P
 	now := time.Now().Unix()
 	cache.CreatedAt = now
 	cache.UsedAt = now
-	if err := insertCache(db, cache); err != nil {
+	if err := h.insertCache(db, cache); err != nil {
 		h.responseJSON(w, r, 500, err)
 		return
 	}
@@ -601,6 +688,12 @@ func (h *Handler) commitCache(cache *Cache) error {
 	cache.Size = written
 	cache.Complete = true
 	cache.UsedAt = time.Now().Unix() // a just-written entry counts as accessed, so it cannot be its own eviction victim
+	if h.s3 != nil {
+		if err := h.s3.publish(cache); err != nil {
+			_ = h.storage.Remove(cache.ID)
+			return err
+		}
+	}
 
 	db, err := h.openDB()
 	if err != nil {
@@ -610,6 +703,9 @@ func (h *Handler) commitCache(cache *Cache) error {
 	if err := db.Update(cache.ID, cache); err != nil {
 		return err
 	}
+	if h.s3 != nil {
+		return nil // lifecycle rules on the bucket bound its size
+	}
 	// A commit is the only thing that grows the store, so the only thing that can push the
 	// volume under the floor.
 	h.evictRepo(db, cache.Repo)
@@ -873,7 +969,13 @@ func findExactCache(db *bolthold.Store, repo, key, version string, complete bool
 	return cache, nil
 }
 
-func insertCache(db *bolthold.Store, cache *Cache) error {
+func (h *Handler) insertCache(db *bolthold.Store, cache *Cache) error {
+	if h.s3 != nil {
+		var id [8]byte
+		_, _ = rand.Read(id[:])
+		cache.ID = binary.BigEndian.Uint64(id[:]) >> 11 // IDs must be unique across runners, and v1 clients parse an ID as a JS number
+		return db.Insert(cache.ID, cache)
+	}
 	return db.Bolt().Update(func(tx *bbolt.Tx) error {
 		if err := db.TxInsert(tx, bolthold.NextSequence(), cache); err != nil {
 			return fmt.Errorf("insert cache: %w", err)
@@ -931,6 +1033,8 @@ const (
 	// a stale reservation.
 	uploadStallTimeout = 5 * time.Minute
 
+	metadataSyncInterval = 30 * time.Second
+
 	defaultMinFreeDisk = 1024 * miB
 )
 
@@ -978,6 +1082,9 @@ func (h *Handler) gcCache() {
 	defer db.Close()
 
 	h.evictIncomplete(db)
+	if h.s3 != nil {
+		return // lifecycle rules on the bucket expire its entries
+	}
 	h.evictExpired(db)
 	h.evictSuperseded(db)
 	h.evictOversized(db)
@@ -1175,7 +1282,11 @@ func totalSize(caches []*Cache) int64 {
 // deleteCache drops an entry and its bytes, reporting whether it went fully. The blob goes
 // first, so a failed unlink leaves the row for the next sweep instead of orphaning bytes.
 func (h *Handler) deleteCache(db *bolthold.Store, cache *Cache) bool {
-	if err := h.storage.Remove(cache.ID); err != nil {
+	storage := h.storage
+	if h.s3 != nil && !cache.Complete {
+		storage = h.s3.Storage // an upload not yet committed has nothing in the bucket
+	}
+	if err := storage.Remove(cache.ID); err != nil {
 		h.logger.Warnf("remove cache blob: %v", err)
 		return false
 	}

act/artifactcache/handler_test.go

@@ -864,7 +864,7 @@ func TestHandler_gcCache(t *testing.T) {
 	db, err := handler.openDB()
 	require.NoError(t, err)
 	for _, c := range cases {
-		require.NoError(t, insertCache(db, c.Cache))
+		require.NoError(t, handler.insertCache(db, c.Cache))
 	}
 	require.NoError(t, db.Close())
 
@@ -1036,7 +1036,7 @@ func TestHandler_SweepKeepsEntryWhenBlobSurvives(t *testing.T) {
 	handler := newTestHandler(t, Policy{Retention: testRetention}, cache)
 
 	// A non-empty directory where the blob belongs makes os.Remove fail on every platform.
-	blob := handler.storage.filename(cache.ID)
+	blob := handler.storage.(*Storage).filename(cache.ID)
 	require.NoError(t, os.MkdirAll(blob, 0o755))
 	require.NoError(t, os.WriteFile(filepath.Join(blob, "held"), []byte("x"), 0o600))
 
@@ -1119,24 +1119,28 @@ func TestHandler_touchCacheSkipsFreshCompletedEntries(t *testing.T) {
 // newTestHandler starts a handler with testToken registered, seeded with entries.
 func newTestHandler(t *testing.T, policy Policy, entries ...*Cache) *Handler {
 	t.Helper()
-	handler, err := StartHandler(Options{
-		Dir:        filepath.Join(t.TempDir(), "artifactcache"),
-		OutboundIP: "127.0.0.1",
-		Policy:     policy,
-	})
-	require.NoError(t, err)
-	t.Cleanup(func() { require.NoError(t, handler.Close()) })
-	handler.RegisterJob(testToken, JobCredential{Repo: testRepo})
+	handler := startTestHandler(t, Options{Policy: policy})
 
 	db, err := handler.openDB()
 	require.NoError(t, err)
 	for _, e := range entries {
-		require.NoError(t, insertCache(db, e))
+		require.NoError(t, handler.insertCache(db, e))
 	}
 	require.NoError(t, db.Close())
 	return handler
 }
 
+func startTestHandler(t *testing.T, opts Options) *Handler {
+	t.Helper()
+	opts.Dir = filepath.Join(t.TempDir(), "artifactcache")
+	opts.OutboundIP = "127.0.0.1"
+	handler, err := StartHandler(opts)
+	require.NoError(t, err)
+	t.Cleanup(func() { require.NoError(t, handler.Close()) })
+	handler.RegisterJob(testToken, JobCredential{Repo: testRepo})
+	return handler
+}
+
 // keptKeys reports which of entries are still in the store.
 func keptKeys(t *testing.T, handler *Handler, entries []*Cache) []string {
 	t.Helper()
@@ -1477,8 +1481,8 @@ func TestHandler_GC_PerRepoDedup(t *testing.T) {
 	stale := time.Now().Add(-inUseGrace - time.Minute).Unix()
 	a := &Cache{Repo: "owner/repoA", Key: key, Version: version, Complete: true, CreatedAt: stale, UsedAt: stale, Size: 1}
 	b := &Cache{Repo: "owner/repoB", Key: key, Version: version, Complete: true, CreatedAt: now, UsedAt: now, Size: 1}
-	require.NoError(t, insertCache(db, a))
-	require.NoError(t, insertCache(db, b))
+	require.NoError(t, handler.insertCache(db, a))
+	require.NoError(t, handler.insertCache(db, b))
 	// Write the backing blobs so the dedup deletion has something to remove.
 	require.NoError(t, handler.storage.Write(a.ID, 0, strings.NewReader("a")))
 	_, err = handler.storage.Commit(a.ID, 1)

act/artifactcache/handler_v2.go

@@ -112,7 +112,7 @@ func (h *Handler) v2CreateCacheEntry(w http.ResponseWriter, r *http.Request, _ h
 		CreatedAt: now,
 		UsedAt:    now,
 	}
-	if err := insertCache(db, cache); err != nil {
+	if err := h.insertCache(db, cache); err != nil {
 		h.twirpError(w, r, twirpInternal, err)
 		return
 	}

act/artifactcache/s3_storage.go

@@ -0,0 +1,326 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package artifactcache
+
+import (
+	"bytes"
+	"cmp"
+	"context"
+	"crypto/hmac"
+	"crypto/sha256"
+	"encoding/json/v2"
+	"encoding/xml"
+	"errors"
+	"fmt"
+	"io"
+	"io/fs"
+	"maps"
+	"net/http"
+	"net/url"
+	"os"
+	"path"
+	"slices"
+	"strconv"
+	"strings"
+	"time"
+)
+
+// S3Options configures the bucket that stores completed cache entries.
+type S3Options struct {
+	Endpoint  string
+	Region    string
+	Bucket    string
+	Prefix    string
+	PathStyle *bool // nil uses virtual-hosted requests on AWS and path-style ones elsewhere
+	AccessKey string
+	SecretKey string
+}
+
+type cacheStorage interface {
+	Exist(uint64) (bool, error)
+	Write(uint64, int64, io.Reader) error
+	WriteBlock(uint64, string, io.Reader) error
+	OrderBlocks(uint64, []string) error
+	Commit(uint64, int64) (int64, error)
+	Serve(http.ResponseWriter, *http.Request, uint64)
+	Remove(uint64) error
+}
+
+const s3PartSize = 5 << 30 // the largest part S3 accepts
+
+// s3Storage stages uploads on disk and keeps completed entries and their metadata in a bucket.
+type s3Storage struct {
+	*Storage
+	client    *http.Client
+	bucketURL *url.URL
+	opts      S3Options
+}
+
+type s3Part struct {
+	PartNumber int
+	ETag       string
+}
+
+func newS3Storage(staging *Storage, opts S3Options) (*s3Storage, error) {
+	endpoint, err := url.Parse(opts.Endpoint)
+	if err != nil || endpoint.Host == "" || (endpoint.Scheme != "http" && endpoint.Scheme != "https") {
+		return nil, fmt.Errorf("invalid S3 endpoint %q", opts.Endpoint)
+	}
+	transport, _ := http.DefaultTransport.(*http.Transport)
+	transport = transport.Clone()
+	transport.ResponseHeaderTimeout = time.Minute // a stalled endpoint must not hold the store forever
+	pathStyle := !strings.HasSuffix(endpoint.Hostname(), ".amazonaws.com")
+	if opts.PathStyle != nil {
+		pathStyle = *opts.PathStyle
+	}
+	if pathStyle {
+		endpoint.Path = "/" + opts.Bucket
+	} else {
+		endpoint.Host = opts.Bucket + "." + endpoint.Host
+		endpoint.Path = "/"
+	}
+	opts.Region = cmp.Or(opts.Region, "us-east-1")
+	return &s3Storage{Storage: staging, client: &http.Client{Transport: transport}, bucketURL: endpoint, opts: opts}, nil
+}
+
+func (s *s3Storage) blobKey(id uint64) string {
+	return path.Join(s.opts.Prefix, "blobs", strconv.FormatUint(id, 10))
+}
+
+func (s *s3Storage) metadataDir() string {
+	return path.Join(s.opts.Prefix, "metadata") + "/"
+}
+
+func (s *s3Storage) metadataKey(id uint64) string {
+	return s.metadataDir() + strconv.FormatUint(id, 10) + ".json"
+}
+
+func (s *s3Storage) url(key string, query url.Values) *url.URL {
+	u := *s.bucketURL
+	u.Path = path.Join(u.Path, key)
+	u.RawPath = strings.NewReplacer("+", "%20", "%2F", "/").Replace(url.QueryEscape(u.Path)) // SigV4 encoding
+	u.RawQuery = strings.ReplaceAll(query.Encode(), "+", "%20")
+	return &u
+}
+
+// sign authorizes req with AWS Signature Version 4, leaving the payload unsigned so bodies can stream.
+// https://docs.aws.amazon.com/AmazonS3/latest/API/sig-v4-header-based-auth.html
+func (s *s3Storage) sign(req *http.Request, now time.Time) {
+	date := now.UTC().Format("20060102T150405Z")
+	req.Header.Set("X-Amz-Date", date)
+	if req.Header.Get("X-Amz-Content-Sha256") == "" {
+		req.Header.Set("X-Amz-Content-Sha256", "UNSIGNED-PAYLOAD")
+	}
+	headers := map[string]string{"host": req.URL.Host}
+	for name, values := range req.Header {
+		if name = strings.ToLower(name); strings.HasPrefix(name, "x-amz-") {
+			headers[name] = strings.Join(values, ",")
+		}
+	}
+	names := slices.Sorted(maps.Keys(headers))
+	canonical := []string{req.Method, req.URL.EscapedPath(), req.URL.RawQuery}
+	for _, name := range names {
+		canonical = append(canonical, name+":"+headers[name])
+	}
+	canonical = append(canonical, "", strings.Join(names, ";"), req.Header.Get("X-Amz-Content-Sha256"))
+	scope := date[:8] + "/" + s.opts.Region + "/s3/aws4_request"
+	signature := []byte("AWS4" + s.opts.SecretKey)
+	for _, data := range []string{date[:8], s.opts.Region, "s3", "aws4_request", fmt.Sprintf("AWS4-HMAC-SHA256\n%s\n%s\n%x", date, scope, sha256.Sum256([]byte(strings.Join(canonical, "\n"))))} { // key derivation, the last round signs
+		mac := hmac.New(sha256.New, signature)
+		_, _ = mac.Write([]byte(data))
+		signature = mac.Sum(nil)
+	}
+	req.Header.Set("Authorization", fmt.Sprintf("AWS4-HMAC-SHA256 Credential=%s/%s,SignedHeaders=%s,Signature=%x", s.opts.AccessKey, scope, strings.Join(names, ";"), signature))
+}
+
+// do returns the response to a signed request when it succeeded, fs.ErrNotExist for a missing key.
+func (s *s3Storage) do(ctx context.Context, method, key string, query url.Values, header http.Header, body io.Reader) (*http.Response, error) {
+	var size int64
+	if sized, ok := body.(interface{ Size() int64 }); ok {
+		size = sized.Size()
+	}
+	if size == 0 {
+		body = nil
+	}
+	req, err := http.NewRequestWithContext(ctx, method, s.url(key, query).String(), body)
+	if err != nil {
+		return nil, err
+	}
+	req.ContentLength = size
+	maps.Copy(req.Header, header)
+	s.sign(req, time.Now())
+	resp, err := s.client.Do(req)
+	if err != nil {
+		return nil, err
+	}
+	if resp.StatusCode < http.StatusMultipleChoices {
+		return resp, nil
+	}
+	defer drain(resp)
+	var failure struct{ Code, Message string }
+	_ = xml.NewDecoder(resp.Body).Decode(&failure)
+	if resp.StatusCode == http.StatusNotFound && (failure.Code == "NoSuchKey" || failure.Code == "" && method == http.MethodHead) {
+		return nil, fs.ErrNotExist
+	}
+	return nil, fmt.Errorf("S3 %s %s: %s", method, key, cmp.Or(strings.TrimSpace(failure.Code+" "+failure.Message), resp.Status))
+}
+
+func (s *s3Storage) call(ctx context.Context, method, key string, query url.Values, body io.Reader, result any) (http.Header, error) {
+	resp, err := s.do(ctx, method, key, query, nil, body)
+	if err != nil {
+		return nil, err
+	}
+	defer drain(resp)
+	if result == nil {
+		return resp.Header, nil
+	}
+	return resp.Header, xml.NewDecoder(resp.Body).Decode(result)
+}
+
+// drain reads a response to its end before closing it, so its connection gets reused.
+func drain(resp *http.Response) {
+	_, _ = io.Copy(io.Discard, resp.Body)
+	_ = resp.Body.Close()
+}
+
+func (s *s3Storage) Exist(id uint64) (bool, error) {
+	_, err := s.call(context.Background(), http.MethodHead, s.blobKey(id), nil, nil, nil)
+	if errors.Is(err, fs.ErrNotExist) {
+		return false, nil
+	}
+	return err == nil, err
+}
+
+func (s *s3Storage) Commit(id uint64, size int64) (int64, error) {
+	written, err := s.Storage.Commit(id, size)
+	if err != nil {
+		return 0, err
+	}
+	defer func() { _ = s.Storage.Remove(id) }()
+	file, err := os.Open(s.filename(id))
+	if err != nil {
+		return 0, err
+	}
+	defer file.Close()
+	if err := s.upload(context.Background(), s.blobKey(id), file, written); err != nil {
+		return 0, fmt.Errorf("upload cache to S3: %w", err)
+	}
+	return written, nil
+}
+
+// upload always goes in parts, the only way S3 takes an object over 5 GiB.
+func (s *s3Storage) upload(ctx context.Context, key string, file io.ReaderAt, size int64) error {
+	var created struct {
+		UploadID string `xml:"UploadId"`
+	}
+	if _, err := s.call(ctx, http.MethodPost, key, url.Values{"uploads": {""}}, nil, &created); err != nil {
+		return err
+	}
+	var completion struct {
+		XMLName xml.Name `xml:"CompleteMultipartUpload"`
+		Parts   []s3Part `xml:"Part"`
+	}
+	for offset := int64(0); offset == 0 || offset < size; offset += s3PartSize {
+		number, length := len(completion.Parts)+1, min(s3PartSize, size-offset)
+		query := url.Values{"partNumber": {strconv.Itoa(number)}, "uploadId": {created.UploadID}}
+		header, err := s.call(ctx, http.MethodPut, key, query, io.NewSectionReader(file, offset, length), nil)
+		if err != nil {
+			return err
+		}
+		completion.Parts = append(completion.Parts, s3Part{number, header.Get("ETag")})
+	}
+	body, err := xml.Marshal(completion)
+	if err != nil {
+		return err
+	}
+	var result struct{ Code, Message string } // a completion can fail after its 200 status
+	if _, err := s.call(ctx, http.MethodPost, key, url.Values{"uploadId": {created.UploadID}}, bytes.NewReader(body), &result); err != nil {
+		return err
+	}
+	if result.Code != "" {
+		return fmt.Errorf("S3 complete upload %s: %s %s", key, result.Code, result.Message)
+	}
+	return nil
+}
+
+func (s *s3Storage) Serve(w http.ResponseWriter, r *http.Request, id uint64) {
+	header := http.Header{}
+	if byteRange := r.Header.Get("Range"); byteRange != "" {
+		header.Set("Range", byteRange)
+	}
+	resp, err := s.do(r.Context(), http.MethodGet, s.blobKey(id), nil, header, nil)
+	if errors.Is(err, fs.ErrNotExist) {
+		http.NotFound(w, r)
+		return
+	} else if err != nil {
+		http.Error(w, "read cache from S3", http.StatusBadGateway)
+		return
+	}
+	defer resp.Body.Close()
+	for _, name := range []string{"Content-Length", "Content-Range"} {
+		if value := resp.Header.Get(name); value != "" {
+			w.Header().Set(name, value)
+		}
+	}
+	w.WriteHeader(resp.StatusCode)
+	_, _ = io.Copy(w, resp.Body)
+}
+
+// Remove unpublishes an entry before deleting its blob, so no runner imports it without its bytes.
+func (s *s3Storage) Remove(id uint64) error {
+	for _, key := range []string{s.metadataKey(id), s.blobKey(id)} {
+		if _, err := s.call(context.Background(), http.MethodDelete, key, nil, nil, nil); err != nil && !errors.Is(err, fs.ErrNotExist) {
+			return fmt.Errorf("remove %s from S3: %w", key, err)
+		}
+	}
+	return s.Storage.Remove(id)
+}
+
+func (s *s3Storage) publish(cache *Cache) error {
+	data, err := json.Marshal(cache)
+	if err != nil {
+		return err
+	}
+	if _, err := s.call(context.Background(), http.MethodPut, s.metadataKey(cache.ID), nil, bytes.NewReader(data), nil); err != nil {
+		return fmt.Errorf("publish cache to S3: %w", err)
+	}
+	return nil
+}
+
+func (s *s3Storage) published(ctx context.Context) (map[uint64]bool, error) {
+	ids := make(map[uint64]bool)
+	query := url.Values{"list-type": {"2"}, "prefix": {s.metadataDir()}}
+	for {
+		var page struct {
+			Contents              []struct{ Key string }
+			IsTruncated           bool
+			NextContinuationToken string
+		}
+		if _, err := s.call(ctx, http.MethodGet, "", query, nil, &page); err != nil {
+			return nil, fmt.Errorf("list S3 cache metadata: %w", err)
+		}
+		for _, object := range page.Contents {
+			if id, err := strconv.ParseUint(strings.TrimSuffix(strings.TrimPrefix(object.Key, s.metadataDir()), ".json"), 10, 64); err == nil {
+				ids[id] = true
+			}
+		}
+		if !page.IsTruncated {
+			return ids, nil
+		}
+		query.Set("continuation-token", page.NextContinuationToken)
+	}
+}
+
+func (s *s3Storage) fetch(ctx context.Context, id uint64) (*Cache, error) {
+	resp, err := s.do(ctx, http.MethodGet, s.metadataKey(id), nil, nil, nil)
+	if err != nil {
+		return nil, err
+	}
+	defer resp.Body.Close()
+	cache := &Cache{}
+	if err := json.UnmarshalRead(resp.Body, cache); err != nil {
+		return nil, fmt.Errorf("read S3 cache metadata %d: %w", id, err)
+	}
+	return cache, nil
+}

act/artifactcache/s3_storage_test.go

@@ -0,0 +1,46 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package artifactcache
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"testing"
+	"time"
+
+	"github.com/johannesboyne/gofakes3"
+	"github.com/johannesboyne/gofakes3/backend/s3mem"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestHandler_S3SharesEntriesAndDropsExpiredOnes(t *testing.T) {
+	backend := s3mem.New()
+	require.NoError(t, backend.CreateBucket("cache"))
+	server := httptest.NewServer(gofakes3.New(backend).Server())
+	t.Cleanup(server.Close)
+	opts := Options{S3: &S3Options{Endpoint: server.URL, Bucket: "cache", Prefix: "runners", AccessKey: "key", SecretKey: "secret"}}
+	saver, restorer := startTestHandler(t, opts), startTestHandler(t, opts)
+	entry := map[string]any{"key": "deps", "version": "v1"}
+
+	finalized, _ := saveV2(t, saver, "deps", "v1", []byte("archive"))
+	require.Equal(t, true, finalized["ok"])
+	require.NoError(t, restorer.importMetadata(t.Context()))
+	found := v2Call(t, restorer, testClient, "GetCacheEntryDownloadURL", entry)
+	require.Equal(t, true, found["ok"])
+	assert.Equal(t, []byte("archive"), getURL(t, found["signed_download_url"].(string)))
+
+	require.NoError(t, backend.ForceDeleteBucket("cache"))
+	require.NoError(t, backend.CreateBucket("cache"))
+	require.NoError(t, restorer.importMetadata(t.Context()))
+	assert.Equal(t, true, v2Call(t, restorer, testClient, "CreateCacheEntry", entry)["ok"])
+}
+
+func TestS3StorageSignsLikeAWS(t *testing.T) {
+	storage := &s3Storage{opts: S3Options{Region: "us-east-1", AccessKey: "AKIAIOSFODNN7EXAMPLE", SecretKey: "wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY"}}
+	req := httptest.NewRequest(http.MethodGet, "https://examplebucket.s3.amazonaws.com/?max-keys=2&prefix=J", nil)
+	req.Header.Set("X-Amz-Content-Sha256", "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855")
+	storage.sign(req, time.Date(2013, 5, 24, 0, 0, 0, 0, time.UTC))
+	assert.Equal(t, "AWS4-HMAC-SHA256 Credential=AKIAIOSFODNN7EXAMPLE/20130524/us-east-1/s3/aws4_request,SignedHeaders=host;x-amz-content-sha256;x-amz-date,Signature=34b48302e7b5fa45bde8084f4b7868a86f0a534bc59db6670ed5711ef69dc6f7", req.Header.Get("Authorization"))
+}