* Add Trino blog operations test * Update test/s3tables/catalog_trino/trino_blog_operations_test.go Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com> * feat: add table bucket path helpers and filer operations - Add table object root and table location mapping directories - Implement ensureDirectory, upsertFile, deleteEntryIfExists helpers - Support table location bucket mapping for S3 access * feat: manage table bucket object roots on creation/deletion - Create .objects directory for table buckets on creation - Clean up table object bucket paths on deletion - Enable S3 operations on table bucket object roots * feat: add table location mapping for Iceberg REST - Track table location bucket mappings when tables are created/updated/deleted - Enable location-based routing for S3 operations on table data * feat: route S3 operations to table bucket object roots - Route table-s3 bucket names to mapped table paths - Route table buckets to object root directories - Support table location bucket mapping lookup * feat: emit table-s3 locations from Iceberg REST - Generate unique table-s3 bucket names with UUID suffix - Store table metadata under table bucket paths - Return table-s3 locations for Trino compatibility * fix: handle missing directories in S3 list operations - Propagate ErrNotFound from ListEntries for non-existent directories - Treat missing directories as empty results for list operations - Fixes Trino non-empty location checks on table creation * test: improve Trino CSV parsing for single-value results - Sanitize Trino output to skip jline warnings - Handle single-value CSV results without header rows - Strip quotes from numeric values in tests * refactor: use bucket path helpers throughout S3 API - Replace direct bucket path operations with helper functions - Leverage centralized table bucket routing logic - Improve maintainability with consistent path resolution * fix: add table bucket cache and improve filer error handling - Cache table bucket lookups to reduce filer overhead on repeated checks - Use filer_pb.CreateEntry and filer_pb.UpdateEntry helpers to check resp.Error - Fix delete order in handler_bucket_get_list_delete: delete table object before directory - Make location mapping errors best-effort: log and continue, don't fail API - Update table location mappings to delete stale prior bucket mappings on update - Add 1-second sleep before timestamp time travel query to ensure timestamps are in past - Fix CSV parsing: examine all lines, not skip first; handle single-value rows * fix: properly handle stale metadata location mapping cleanup - Capture oldMetadataLocation before mutation in handleUpdateTable - Update updateTableLocationMapping to accept both old and new locations - Use passed-in oldMetadataLocation to detect location changes - Delete stale mapping only when location actually changes - Pass empty string for oldLocation in handleCreateTable (new tables have no prior mapping) - Improve logging to show old -> new location transitions * refactor: cleanup imports and cache design - Remove unused 'sync' import from bucket_paths.go - Use filer_pb.UpdateEntry helper in setExtendedAttribute and deleteExtendedAttribute for consistent error handling - Add dedicated tableBucketCache map[string]bool to BucketRegistry instead of mixing concerns with metadataCache - Improve cache separation: table buckets cache is now separate from bucket metadata cache * fix: improve cache invalidation and add transient error handling Cache invalidation (critical fix): - Add tableLocationCache to BucketRegistry for location mapping lookups - Clear tableBucketCache and tableLocationCache in RemoveBucketMetadata - Prevents stale cache entries when buckets are deleted/recreated Transient error handling: - Only cache table bucket lookups when conclusive (found or ErrNotFound) - Skip caching on transient errors (network, permission, etc) - Prevents marking real table buckets as non-table due to transient failures Performance optimization: - Cache tableLocationDir results to avoid repeated filer RPCs on hot paths - tableLocationDir now checks cache before making expensive filer lookups - Cache stores empty string for 'not found' to avoid redundant lookups Code clarity: - Add comment to deleteDirectory explaining DeleteEntry response lacks Error field * go fmt * fix: mirror transient error handling in tableLocationDir and optimize bucketDir Transient error handling: - tableLocationDir now only caches definitive results - Mirrors isTableBucket behavior to prevent treating transient errors as permanent misses - Improves reliability on flaky systems or during recovery Performance optimization: - bucketDir avoids redundant isTableBucket call via bucketRoot - Directly use s3a.option.BucketsPath for regular buckets - Saves one cache lookup for every non-table bucket operation * fix: revert bucketDir optimization to preserve bucketRoot logic The optimization to directly use BucketsPath bypassed bucketRoot's logic and caused issues with S3 list operations on delimiter+prefix cases. Revert to using path.Join(s3a.bucketRoot(bucket), bucket) which properly handles all bucket types and ensures consistent path resolution across the codebase. The slight performance cost of an extra cache lookup is worth the correctness and consistency benefits. * feat: move table buckets under /buckets Add a table-bucket marker attribute, reuse bucket metadata cache for table bucket detection, and update list/validation/UI/test paths to treat table buckets as /buckets entries. * Fix S3 Tables code review issues - handler_bucket_create.go: Fix bucket existence check to properly validate entryResp.Entry before setting s3BucketExists flag (nil Entry should not indicate existing bucket) - bucket_paths.go: Add clarifying comment to bucketRoot() explaining unified buckets root path for all bucket types - file_browser_data.go: Optimize by extracting table bucket check early to avoid redundant WithFilerClient call * Fix list prefix delimiter handling * Handle list errors conservatively * Fix Trino FOR TIMESTAMP query - use past timestamp Iceberg requires the timestamp to be strictly in the past. Use current_timestamp - interval '1' second instead of current_timestamp. --------- Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
185 lines
6.0 KiB
Go
185 lines
6.0 KiB
Go
package s3api
|
||
|
||
import (
|
||
"crypto/rand"
|
||
"encoding/hex"
|
||
"fmt"
|
||
"math"
|
||
"strconv"
|
||
"time"
|
||
|
||
"github.com/seaweedfs/seaweedfs/weed/glog"
|
||
s3_constants "github.com/seaweedfs/seaweedfs/weed/s3api/s3_constants"
|
||
)
|
||
|
||
// Version ID format constants
|
||
// New format uses inverted timestamps so newer versions sort first lexicographically
|
||
// Old format used raw timestamps where older versions sorted first
|
||
const (
|
||
// Threshold to distinguish old vs new format version IDs
|
||
// Around year 2024-2025:
|
||
// - Old format (raw ns): ~1.7×10¹⁸ ≈ 0x17... (BELOW threshold)
|
||
// - New format (MaxInt64 - ns): ~7.5×10¹⁸ ≈ 0x68... (ABOVE threshold)
|
||
// We use 0x4000000000000000 (~4.6×10¹⁸) as threshold
|
||
versionIdFormatThreshold = 0x4000000000000000
|
||
)
|
||
|
||
// generateVersionId creates a unique version ID
|
||
// If useInvertedFormat is true, uses inverted timestamps so newer versions sort first
|
||
// If false, uses raw timestamps (old format) for backward compatibility
|
||
func generateVersionId(useInvertedFormat bool) string {
|
||
now := time.Now().UnixNano()
|
||
var timestampHex string
|
||
|
||
if useInvertedFormat {
|
||
// INVERTED timestamp: newer versions have SMALLER values
|
||
// This makes lexicographic sorting return newest versions first
|
||
invertedTimestamp := math.MaxInt64 - now
|
||
timestampHex = fmt.Sprintf("%016x", invertedTimestamp)
|
||
} else {
|
||
// Raw timestamp: older versions have SMALLER values (old format)
|
||
timestampHex = fmt.Sprintf("%016x", now)
|
||
}
|
||
|
||
// Generate random 8 bytes for uniqueness (last 16 chars of version ID)
|
||
randBytes := make([]byte, 8)
|
||
if _, err := rand.Read(randBytes); err != nil {
|
||
glog.Errorf("Failed to generate random bytes for version ID: %v", err)
|
||
// Fallback to timestamp-only if random generation fails
|
||
return timestampHex + "0000000000000000"
|
||
}
|
||
|
||
// Combine timestamp (16 chars) + random (16 chars) = 32 chars total
|
||
randomHex := hex.EncodeToString(randBytes)
|
||
return timestampHex + randomHex
|
||
}
|
||
|
||
// isNewFormatVersionId returns true if the version ID uses the new inverted timestamp format
|
||
func isNewFormatVersionId(versionId string) bool {
|
||
if len(versionId) < 16 || versionId == "null" {
|
||
return false
|
||
}
|
||
// Parse the first 16 hex chars as the timestamp portion
|
||
timestampPart, err := strconv.ParseUint(versionId[:16], 16, 64)
|
||
if err != nil {
|
||
return false
|
||
}
|
||
// New format has inverted timestamps (MaxInt64 - ns), which are ABOVE the threshold (~0x68...)
|
||
// Old format has raw timestamps, which are BELOW the threshold (~0x17...)
|
||
return timestampPart > versionIdFormatThreshold
|
||
}
|
||
|
||
// getVersionTimestamp extracts the actual timestamp from a version ID,
|
||
// handling both old (raw) and new (inverted) formats
|
||
func getVersionTimestamp(versionId string) int64 {
|
||
if len(versionId) < 16 || versionId == "null" {
|
||
return 0
|
||
}
|
||
timestampPart, err := strconv.ParseUint(versionId[:16], 16, 64)
|
||
if err != nil {
|
||
return 0
|
||
}
|
||
if timestampPart > versionIdFormatThreshold {
|
||
// New format: inverted timestamp (above threshold), convert back
|
||
return int64(math.MaxInt64 - timestampPart)
|
||
}
|
||
// Validate old format timestamp is within int64 range
|
||
if timestampPart > math.MaxInt64 {
|
||
return 0
|
||
}
|
||
// Old format: raw timestamp (below threshold)
|
||
return int64(timestampPart)
|
||
}
|
||
|
||
// compareVersionIds compares two version IDs for sorting (newest first)
|
||
// Returns: negative if a is newer, positive if b is newer, 0 if equal
|
||
// Handles both old and new format version IDs
|
||
func compareVersionIds(a, b string) int {
|
||
if a == b {
|
||
return 0
|
||
}
|
||
if a == "null" {
|
||
return 1 // null versions sort last
|
||
}
|
||
if b == "null" {
|
||
return -1
|
||
}
|
||
|
||
aIsNew := isNewFormatVersionId(a)
|
||
bIsNew := isNewFormatVersionId(b)
|
||
|
||
if aIsNew == bIsNew {
|
||
// Same format - compare lexicographically
|
||
// For new format: smaller value = newer (correct)
|
||
// For old format: smaller value = older (need to invert)
|
||
if aIsNew {
|
||
// New format: lexicographic order is correct (smaller = newer)
|
||
if a < b {
|
||
return -1
|
||
}
|
||
return 1
|
||
} else {
|
||
// Old format: lexicographic order is inverted (smaller = older)
|
||
if a < b {
|
||
return 1
|
||
}
|
||
return -1
|
||
}
|
||
}
|
||
|
||
// Mixed formats - compare by actual timestamp
|
||
aTime := getVersionTimestamp(a)
|
||
bTime := getVersionTimestamp(b)
|
||
if aTime > bTime {
|
||
return -1 // a is newer
|
||
}
|
||
if aTime < bTime {
|
||
return 1 // b is newer
|
||
}
|
||
return 0
|
||
}
|
||
|
||
// getVersionedObjectDir returns the directory path for storing object versions
|
||
func (s3a *S3ApiServer) getVersionedObjectDir(bucket, object string) string {
|
||
return s3a.bucketDir(bucket) + "/" + object + s3_constants.VersionsFolder
|
||
}
|
||
|
||
// getVersionFileName returns the filename for a specific version
|
||
func (s3a *S3ApiServer) getVersionFileName(versionId string) string {
|
||
return fmt.Sprintf("v_%s", versionId)
|
||
}
|
||
|
||
// getVersionIdFormat checks the .versions directory to determine which version ID format to use.
|
||
// Returns true if inverted format (new format) should be used.
|
||
// For new .versions directories, returns true (use new format).
|
||
// For existing directories, infers format from the latest version ID.
|
||
func (s3a *S3ApiServer) getVersionIdFormat(bucket, object string) bool {
|
||
bucketDir := s3a.bucketDir(bucket)
|
||
versionsPath := object + s3_constants.VersionsFolder
|
||
|
||
// Try to get the .versions directory entry
|
||
versionsEntry, err := s3a.getEntry(bucketDir, versionsPath)
|
||
if err != nil {
|
||
// .versions directory doesn't exist yet - use new format
|
||
return true
|
||
}
|
||
|
||
// Infer format from the latest version ID stored in metadata
|
||
if versionsEntry.Extended != nil {
|
||
if latestVersionId, exists := versionsEntry.Extended[s3_constants.ExtLatestVersionIdKey]; exists {
|
||
return isNewFormatVersionId(string(latestVersionId))
|
||
}
|
||
}
|
||
|
||
// No latest version metadata - this is likely a new or empty directory
|
||
// Use new format
|
||
return true
|
||
}
|
||
|
||
// generateVersionIdForObject generates a version ID using the appropriate format for the object.
|
||
// For new objects, uses inverted format. For existing versioned objects, uses their existing format.
|
||
func (s3a *S3ApiServer) generateVersionIdForObject(bucket, object string) string {
|
||
useInvertedFormat := s3a.getVersionIdFormat(bucket, object)
|
||
return generateVersionId(useInvertedFormat)
|
||
}
|