Improve object storage performance via caching

This commit is contained in:
2025-12-22 17:03:33 +08:00
parent aab9ef696a
commit f5451c162b
5 changed files with 542 additions and 311 deletions

View File

@@ -128,11 +128,14 @@ class ObjectStorage:
BUCKET_VERSIONS_DIR = "versions"
MULTIPART_MANIFEST = "manifest.json"
BUCKET_CONFIG_FILE = ".bucket.json"
KEY_INDEX_CACHE_TTL = 30 # seconds - longer TTL for better browsing performance
def __init__(self, root: Path) -> None:
self.root = Path(root)
self.root.mkdir(parents=True, exist_ok=True)
self._ensure_system_roots()
# In-memory object metadata cache: bucket_id -> (dict[key -> ObjectMeta], timestamp)
self._object_cache: Dict[str, tuple[Dict[str, ObjectMeta], float]] = {}
def list_buckets(self) -> List[BucketMeta]:
buckets: List[BucketMeta] = []
@@ -274,32 +277,26 @@ class ObjectStorage:
raise StorageError("Bucket does not exist")
bucket_id = bucket_path.name
# Collect all matching object keys first (lightweight - just paths)
all_keys: List[str] = []
for path in bucket_path.rglob("*"):
if path.is_file():
rel = path.relative_to(bucket_path)
if rel.parts and rel.parts[0] in self.INTERNAL_FOLDERS:
continue
key = str(rel.as_posix())
if prefix and not key.startswith(prefix):
continue
all_keys.append(key)
# Use cached object metadata for fast listing
object_cache = self._get_object_cache(bucket_id, bucket_path)
# Get sorted keys
all_keys = sorted(object_cache.keys())
# Apply prefix filter if specified
if prefix:
all_keys = [k for k in all_keys if k.startswith(prefix)]
all_keys.sort()
total_count = len(all_keys)
# Handle continuation token (the key to start after)
start_index = 0
if continuation_token:
try:
# continuation_token is the last key from previous page
for i, key in enumerate(all_keys):
if key > continuation_token:
start_index = i
break
else:
# Token is past all keys
# Binary search for efficiency on large lists
import bisect
start_index = bisect.bisect_right(all_keys, continuation_token)
if start_index >= total_count:
return ListObjectsResult(
objects=[],
is_truncated=False,
@@ -314,27 +311,12 @@ class ObjectStorage:
keys_slice = all_keys[start_index:end_index]
is_truncated = end_index < total_count
# Now load full metadata only for the objects we're returning
# Build result from cached metadata (no file I/O!)
objects: List[ObjectMeta] = []
for key in keys_slice:
safe_key = self._sanitize_object_key(key)
path = bucket_path / safe_key
if not path.exists():
continue # Object may have been deleted
try:
stat = path.stat()
metadata = self._read_metadata(bucket_id, safe_key)
objects.append(
ObjectMeta(
key=key,
size=stat.st_size,
last_modified=datetime.fromtimestamp(stat.st_mtime),
etag=self._compute_etag(path),
metadata=metadata or None,
)
)
except OSError:
continue # File may have been deleted during iteration
obj = object_cache.get(key)
if obj:
objects.append(obj)
next_token = keys_slice[-1] if is_truncated and keys_slice else None
@@ -416,18 +398,21 @@ class ObjectStorage:
pass
stat = destination.stat()
if metadata:
self._write_metadata(bucket_id, safe_key, metadata)
else:
self._delete_metadata(bucket_id, safe_key)
etag = checksum.hexdigest()
# Always store internal metadata (etag, size) alongside user metadata
internal_meta = {"__etag__": etag, "__size__": str(stat.st_size)}
combined_meta = {**internal_meta, **(metadata or {})}
self._write_metadata(bucket_id, safe_key, combined_meta)
self._invalidate_bucket_stats_cache(bucket_id)
self._invalidate_object_cache(bucket_id)
return ObjectMeta(
key=safe_key.as_posix(),
size=stat.st_size,
last_modified=datetime.fromtimestamp(stat.st_mtime),
etag=checksum.hexdigest(),
etag=etag,
metadata=metadata,
)
@@ -479,6 +464,7 @@ class ObjectStorage:
self._delete_metadata(bucket_id, rel)
self._invalidate_bucket_stats_cache(bucket_id)
self._invalidate_object_cache(bucket_id)
self._cleanup_empty_parents(path, bucket_path)
def purge_object(self, bucket_name: str, object_key: str) -> None:
@@ -501,6 +487,7 @@ class ObjectStorage:
# Invalidate bucket stats cache
self._invalidate_bucket_stats_cache(bucket_id)
self._invalidate_object_cache(bucket_id)
self._cleanup_empty_parents(target, bucket_path)
def is_versioning_enabled(self, bucket_name: str) -> bool:
@@ -1163,6 +1150,187 @@ class ObjectStorage:
def _legacy_multipart_dir(self, bucket_name: str, upload_id: str) -> Path:
return self._legacy_multipart_bucket_root(bucket_name) / upload_id
def _fast_list_keys(self, bucket_path: Path) -> List[str]:
"""Fast directory walk using os.scandir instead of pathlib.rglob.
This is significantly faster for large directories (10K+ files).
Returns just the keys (for backward compatibility).
"""
return list(self._build_object_cache(bucket_path).keys())
def _build_object_cache(self, bucket_path: Path) -> Dict[str, ObjectMeta]:
"""Build a complete object metadata cache for a bucket.
Uses os.scandir for fast directory walking and a persistent etag index.
"""
from concurrent.futures import ThreadPoolExecutor
bucket_id = bucket_path.name
objects: Dict[str, ObjectMeta] = {}
bucket_str = str(bucket_path)
bucket_len = len(bucket_str) + 1 # +1 for the separator
# Try to load persisted etag index first (single file read vs thousands)
etag_index_path = self._system_bucket_root(bucket_id) / "etag_index.json"
meta_cache: Dict[str, str] = {}
index_mtime: float = 0
if etag_index_path.exists():
try:
index_mtime = etag_index_path.stat().st_mtime
with open(etag_index_path, 'r', encoding='utf-8') as f:
meta_cache = json.load(f)
except (OSError, json.JSONDecodeError):
meta_cache = {}
# Check if we need to rebuild the index
meta_root = self._bucket_meta_root(bucket_id)
needs_rebuild = False
if meta_root.exists() and index_mtime > 0:
# Quick check: if any meta file is newer than index, rebuild
def check_newer(dir_path: str) -> bool:
try:
with os.scandir(dir_path) as it:
for entry in it:
if entry.is_dir(follow_symlinks=False):
if check_newer(entry.path):
return True
elif entry.is_file(follow_symlinks=False) and entry.name.endswith('.meta.json'):
if entry.stat().st_mtime > index_mtime:
return True
except OSError:
pass
return False
needs_rebuild = check_newer(str(meta_root))
elif not meta_cache:
needs_rebuild = True
if needs_rebuild and meta_root.exists():
meta_str = str(meta_root)
meta_len = len(meta_str) + 1
meta_files: list[tuple[str, str]] = []
# Collect all metadata file paths
def collect_meta_files(dir_path: str) -> None:
try:
with os.scandir(dir_path) as it:
for entry in it:
if entry.is_dir(follow_symlinks=False):
collect_meta_files(entry.path)
elif entry.is_file(follow_symlinks=False) and entry.name.endswith('.meta.json'):
rel = entry.path[meta_len:]
key = rel[:-10].replace(os.sep, '/')
meta_files.append((key, entry.path))
except OSError:
pass
collect_meta_files(meta_str)
# Parallel read of metadata files - only extract __etag__
def read_meta_file(item: tuple[str, str]) -> tuple[str, str | None]:
key, path = item
try:
with open(path, 'rb') as f:
content = f.read()
etag_marker = b'"__etag__"'
idx = content.find(etag_marker)
if idx != -1:
start = content.find(b'"', idx + len(etag_marker) + 1)
if start != -1:
end = content.find(b'"', start + 1)
if end != -1:
return key, content[start+1:end].decode('utf-8')
return key, None
except (OSError, UnicodeDecodeError):
return key, None
if meta_files:
meta_cache = {}
with ThreadPoolExecutor(max_workers=min(64, len(meta_files))) as executor:
for key, etag in executor.map(read_meta_file, meta_files):
if etag:
meta_cache[key] = etag
# Persist the index for next time
try:
etag_index_path.parent.mkdir(parents=True, exist_ok=True)
with open(etag_index_path, 'w', encoding='utf-8') as f:
json.dump(meta_cache, f)
except OSError:
pass
# Now scan objects and use cached etags
def scan_dir(dir_path: str) -> None:
try:
with os.scandir(dir_path) as it:
for entry in it:
if entry.is_dir(follow_symlinks=False):
# Skip internal folders
rel_start = entry.path[bucket_len:].split(os.sep)[0] if len(entry.path) > bucket_len else entry.name
if rel_start in self.INTERNAL_FOLDERS:
continue
scan_dir(entry.path)
elif entry.is_file(follow_symlinks=False):
# Get relative path and convert to POSIX
rel = entry.path[bucket_len:]
# Check if in internal folder
first_part = rel.split(os.sep)[0] if os.sep in rel else rel
if first_part in self.INTERNAL_FOLDERS:
continue
key = rel.replace(os.sep, '/')
try:
# Use entry.stat() which is cached from scandir
stat = entry.stat()
# Get etag from cache (now just a string, not dict)
etag = meta_cache.get(key)
# Use placeholder for legacy objects without stored etag
if not etag:
etag = f'"{stat.st_size}-{int(stat.st_mtime)}"'
objects[key] = ObjectMeta(
key=key,
size=stat.st_size,
last_modified=datetime.fromtimestamp(stat.st_mtime),
etag=etag,
metadata=None, # Don't include user metadata in listing
)
except OSError:
pass
except OSError:
pass
scan_dir(bucket_str)
return objects
def _get_object_cache(self, bucket_id: str, bucket_path: Path) -> Dict[str, ObjectMeta]:
"""Get cached object metadata for a bucket, refreshing if stale."""
now = time.time()
cached = self._object_cache.get(bucket_id)
if cached:
objects, timestamp = cached
if now - timestamp < self.KEY_INDEX_CACHE_TTL:
return objects
# Rebuild cache
objects = self._build_object_cache(bucket_path)
self._object_cache[bucket_id] = (objects, now)
return objects
def _invalidate_object_cache(self, bucket_id: str) -> None:
"""Invalidate the object cache and etag index for a bucket."""
self._object_cache.pop(bucket_id, None)
# Also invalidate persisted etag index
etag_index_path = self._system_bucket_root(bucket_id) / "etag_index.json"
try:
etag_index_path.unlink(missing_ok=True)
except OSError:
pass
def _ensure_system_roots(self) -> None:
for path in (
self._system_root_path(),