CDB64 File Format Specification

February 11, 2026 · View on GitHub

CDB64 is a 64-bit variant of D. J. Bernstein's Constant Database (CDB) format. It provides O(1) key-value lookups with minimal overhead, designed for static datasets that are written once and read many times.

Table of Contents

Overview

CDB64 extends the original CDB format to support files larger than 4GB by using 64-bit file offsets instead of 32-bit. This is necessary for large historical indexes that can exceed the 4GB limit.

Key Characteristics

  • Immutable: Files are written once and never modified
  • O(1) Lookups: Constant-time key lookups via hash tables
  • Compact: Minimal overhead per record
  • Simple: Easy to implement in any language
  • Portable: Little-endian byte order, no alignment requirements

File Structure

A CDB64 file consists of three sections in order:

$\text{text} +------------------+ | \text{Header} | 4096 \text{bytes} (256 \times 16-\text{byte} \text{pointers}) +------------------+ | \text{Records} | \text{Variable} \text{length} +------------------+ | \text{Hash} \text{Tables} | \text{Variable} \text{length} (256 \text{tables}) +------------------+ $

Header (4096 bytes)

The header contains 256 table pointers, one for each possible value of the low 8 bits of a hash. Each pointer is 16 bytes:

OffsetSizeTypeDescription
08uint64_leTable position (byte offset from start of file)
88uint64_leTable length (number of slots, not bytes)

Total header size: 256 × 16 = 4096 bytes

Records

Records are stored sequentially starting at byte offset 4096. Each record has the format:

OffsetSizeTypeDescription
08uint64_leKey length in bytes
88uint64_leValue length in bytes
16key_lengthbytesKey data
16 + key_lengthvalue_lengthbytesValue data

Records have no padding or alignment requirements.

Hash Tables

After all records, the file contains 256 hash tables. Each table has a variable number of slots (stored in the header). Each slot is 16 bytes:

OffsetSizeTypeDescription
08uint64_leFull 64-bit hash of the key
88uint64_leRecord position (0 if slot is empty)

The number of slots in each table is always 2× the number of records that hash to that table (for efficient linear probing).

Hash Function

CDB64 uses the DJB hash function (same as original CDB), computed with 64-bit arithmetic:

hash = 5381
for each byte in key:
    hash = ((hash << 5) + hash) ^ byte
    hash = hash & 0xFFFFFFFFFFFFFFFF  // Keep as 64-bit unsigned

The hash is used in two ways:

  1. hash % 256 → Selects which of the 256 tables to use
  2. (hash / 256) % table_length → Starting slot within that table

Lookup Algorithm

To look up a key:

  1. Compute hash = djb_hash(key) (64-bit)
  2. Select table: table_index = hash % 256
  3. Read table pointer from header at offset table_index × 16
  4. If table length is 0, key is not present
  5. Compute starting slot: slot = (hash / 256) % table_length
  6. Linear probe:
    • Read slot at table_position + slot × 16
    • If slot position is 0, key is not present
    • If slot hash matches and key matches, return value
    • Otherwise: slot = (slot + 1) % table_length, repeat

Writing Algorithm

To create a CDB64 file:

  1. Skip header (write 4096 zero bytes or seek past it)
  2. For each key-value pair:
    • Write record (key_len, value_len, key, value)
    • Remember: { hash, position } for each record
    • Group records by hash % 256
  3. For each of the 256 tables:
    • Table length = 2 × number of records in this table
    • Create slots array initialized to { hash: 0, position: 0 }
    • For each record in this table:
      • Starting slot = (hash / 256) % table_length
      • Linear probe to find empty slot (position == 0)
      • Store { hash, position } in slot
    • Write all slots
    • Record table position and length
  4. Seek to beginning, write header with table pointers

Differences from Original CDB

AspectOriginal CDBCDB64
Header pointer size8 bytes (4+4)16 bytes (8+8)
Header total size2048 bytes4096 bytes
Record header size8 bytes (4+4)16 bytes (8+8)
Hash table slot size8 bytes (4+4)16 bytes (8+8)
Hash value32-bit64-bit
Key/value length fields32-bit64-bit
Position fields32-bit64-bit
Maximum file size4 GB16 EB (practical: unlimited)

The hash function algorithm is identical (DJB hash), but computed with 64-bit arithmetic.

Root TX Index Value Format

For the AR.IO Gateway's root TX index, CDB64 keys and values have specific formats.

Key Format

Keys are 32-byte data item IDs (binary, not base64-encoded):

Key: <32 bytes> - Raw data item ID

Value Format

Values are MessagePack-encoded objects with short keys for compactness. Four formats are supported to handle both legacy data and nested bundle hierarchies.

Legacy Formats

Simple Format

Used when only the root transaction ID is known:

{
  r: <Buffer 32 bytes>  // Root transaction ID (binary)
}

Complete Format

Used when offset information is available:

{
  r: <Buffer 32 bytes>,  // Root transaction ID (binary)
  i: <integer>,          // Root data item offset (byte offset of data item header)
  d: <integer>           // Root data offset (byte offset of data payload)
}

Path Formats

Path formats are used for nested bundles (bundles containing bundles). The path array provides the traversal route from the L1 root transaction through intermediate bundles to the immediate parent bundle.

Path Format

Used when the bundle traversal path is known but offsets are not:

{
  p: [<Buffer 32 bytes>, ...]  // Array of bundle IDs from root to parent
}

Path Complete Format

Used when both path and offset information are available:

{
  p: [<Buffer 32 bytes>, ...],  // Array of bundle IDs from root to parent
  i: <integer>,                  // Root data item offset
  d: <integer>                   // Root data offset
}

Path Structure

The path array contains transaction/data item IDs representing the bundle hierarchy:

  • path[0] is always the L1 root transaction ID
  • path[1..n-1] are intermediate nested bundle IDs (if any)
  • path[n-1] (last element) is the immediate parent bundle containing the data item
  • The data item ID itself is NOT included in the path

Example path for a deeply nested data item:

Root TX → Bundle A → Bundle B → Data Item
path = [RootTxId, BundleAId, BundleBId]

For path formats, the root TX ID is derived from path[0], eliminating the need for a separate r field.

Field Mapping

MessagePack KeyFull NameDescription
rrootTxId32-byte root transaction ID (legacy formats only)
ppathArray of 32-byte bundle IDs [root, ..., parent]
irootDataItemOffsetByte offset of nested data item within root TX
drootDataOffsetByte offset of data payload within root TX

These offsets correspond to the HTTP headers:

  • i → X-AR-IO-Root-Data-Item-Offset
  • d → X-AR-IO-Root-Data-Offset

Maximum Nesting Depth

The path array is limited to a maximum of 10 elements (MAX_BUNDLE_NESTING_DEPTH), which supports bundle nesting up to 9 levels deep (root TX + 9 nested bundles).

Examples

Legacy Complete Format

For a data item abc123... nested directly in root TX xyz789... at offset 1024 with data at offset 1536:

Key (32 bytes):

abc123... (raw binary data item ID)

Value (MessagePack encoded):

{
  r: Buffer<xyz789...>,  // 32 bytes
  i: 1024,
  d: 1536
}

Encoded size: ~40-45 bytes depending on offset values

Path Complete Format

For a data item def456... nested inside Bundle B (bbb...) which is inside Bundle A (aaa...) which is inside root TX xyz789...:

Key (32 bytes):

def456... (raw binary data item ID)

Value (MessagePack encoded):

{
  p: [
    Buffer<xyz789...>,  // Root TX ID (32 bytes)
    Buffer<aaa...>,     // Bundle A ID (32 bytes)
    Buffer<bbb...>      // Bundle B ID - immediate parent (32 bytes)
  ],
  i: 5000,
  d: 5512
}

Encoded size: ~110-115 bytes for 3-element path with offsets

Implementation Notes

File Creation

CDB64 files should be created atomically:

  1. Write to a temporary file (e.g., output.cdb.tmp.{pid})
  2. Rename to final path after completion

This ensures readers never see partial files.

Concurrency

  • Multiple readers can safely read the same file concurrently
  • Writers should use exclusive access during creation
  • Files are immutable after creation - no locking needed for reads

Memory Usage

The reader only needs to keep the 4096-byte header in memory. All lookups are done via direct file seeks, making it suitable for very large files.

Error Handling

  • Invalid files can be detected by checking if table positions are within file bounds
  • Corrupted records can be detected by checking if key/value lengths are reasonable
  • Hash collisions are handled correctly by the linear probing algorithm

File Size Limits

While CDB64 uses 64-bit offsets supporting files up to 16 exabytes theoretically, practical limits apply:

  • JavaScript/Node.js: ~8 PiB (~9 PB) (Number.MAX_SAFE_INTEGER = 2532^{53} - 1 bytes)
  • Rust cdb64-rs: Full 64-bit support

The AR.IO Gateway implementation will throw an error for file positions exceeding the safe integer limit.

Partitioned CDB64 Index Format

For very large indexes, CDB64 files can be partitioned by key prefix into up to 256 separate files. This enables:

  • Manageable file sizes: Each partition contains only keys with a specific first byte
  • Parallel I/O: Different partitions can be accessed concurrently
  • Lazy loading: Only open partitions that are actually accessed
  • Flexible storage: Partitions can be stored locally, on HTTP servers, or on Arweave

Directory Structure

A partitioned index consists of a directory containing:

index/
  manifest.json    # Index manifest with partition metadata
  00.cdb           # Records with keys starting 0x00
  01.cdb           # Records with keys starting 0x01
  ...
  ff.cdb           # Records with keys starting 0xff

Not all 256 partition files need to exist - only partitions that contain records are created.

Manifest Format

The manifest.json file describes the partitioned index:

{
  "version": 1,
  "createdAt": "2025-01-15T12:00:00.000Z",
  "totalRecords": 1000000,
  "partitions": [
    {
      "prefix": "00",
      "location": { "type": "file", "filename": "00.cdb" },
      "recordCount": 3921,
      "size": 245760,
      "sha256": "abc123..."
    },
    {
      "prefix": "01",
      "location": { "type": "file", "filename": "01.cdb" },
      "recordCount": 3847,
      "size": 241664
    }
  ],
  "metadata": {
    "source": "custom metadata"
  }
}

Manifest Fields

FieldTypeRequiredDescription
versionintegerYesManifest format version (currently 1)
createdAtstringYesISO 8601 creation timestamp
totalRecordsintegerYesTotal records across all partitions
partitionsarrayYesList of partition descriptors
metadataobjectNoOptional custom metadata

Partition Fields

FieldTypeRequiredDescription
prefixstringYesTwo-character lowercase hex prefix ("00" - "ff")
locationobjectYesLocation descriptor (see below)
recordCountintegerYesNumber of records in this partition
sizeintegerYesFile size in bytes
sha256stringNoSHA-256 hash for integrity verification

Location Types

Partitions support flexible storage locations:

File Location

Partition stored as a local file:

{ "type": "file", "filename": "00.cdb" }

HTTP Location

Partition accessible via HTTP(S):

{ "type": "http", "url": "https://example.com/index/00.cdb" }

Arweave ID Location

Partition accessed by Arweave ID (transaction or data item):

{ "type": "arweave-id", "id": "abc123..." }

Arweave Byte Range Location

Partition stored within an Arweave transaction, accessed via byte-range:

{
  "type": "arweave-byte-range",
  "rootTxId": "abc123...",
  "dataOffsetInRootTx": 1024,
  "dataItemId": "xyz789..."
}
FieldRequiredDescription
rootTxIdYesL1 transaction ID containing the data
dataOffsetInRootTxYesByte offset where partition data begins
dataItemIdNoOriginal data item ID (for fallback/debugging)

Note: Partition size comes from partition.size, not the location.

Partitioning Scheme

Records are partitioned based on the first byte of the 32-byte key:

  • Key prefix 0x00 → 00.cdb
  • Key prefix 0x01 → 01.cdb
  • ...
  • Key prefix 0xff → ff.cdb

This provides uniform distribution for random keys (like transaction IDs).

Lookup Algorithm

To look up a key in a partitioned index:

  1. Read the first byte of the key to determine prefix
  2. Find the partition with matching prefix in the manifest
  3. If no partition exists for that prefix, key is not present
  4. Open/read the partition CDB64 file
  5. Look up the key using standard CDB64 lookup

Reader Implementation

The partitioned reader (PartitionedCdb64Reader) features:

  • Lazy partition opening: Partitions are only opened when first accessed
  • Graceful degradation: Missing partition files return undefined rather than throwing
  • Multiple location types: Supports file, HTTP, and Arweave sources
  • Caching for remote sources: HTTP and Arweave partitions use byte-range caching

Writer Implementation

The partitioned writer (PartitionedCdb64Writer) features:

  • Lazy partition creation: Partition files are only created when records arrive
  • Atomic directory creation: Writes to temp directory, then renames atomically
  • Automatic manifest generation: Creates manifest.json with all partition metadata