Migration / Replication record layout

This document describes the on-the-wire byte layout of the records that cluster key migration and replication ship in the send buffer, in both the non-chunked (whole record in one buffer) and chunked (record too large for one buffer) forms.

Keep this document in sync with the code. If any of the following change, update the layouts and diagrams below:

• libs/storage/Tsavorite/cs/src/core/Allocator/DiskLogRecord.cs — Serialize / Deserialize, the chunked helpers (GetChunkedRecordInlineSize, CompleteDeserializeChunkedRecord), and SerializeInlinePortionForMigration.
• libs/storage/Tsavorite/cs/src/core/Allocator/RecordDataHeader.cs — the record data header (indicator word).
• libs/client/ClientSession/GarnetClientSessionIncremental.cs — MigrationRecordSpanType, TryWriteRecordSpan, TryWriteChunkedRecordSpan, and the send-buffer framing.
• libs/cluster/Session/ChunkedRecordReassembler.cs — the receive-side reassembler.
• libs/cluster/Session/RespClusterMigrateCommands.cs / RespClusterReplicationCommands.cs — the receive paths (NetworkClusterMigrate / NetworkClusterSync, CompleteChunkedRecordReassembly).
• libs/server/MigrationChunkWriterAccumulator.cs / libs/server/Storage/Functions/UnifiedStore/ReadMethods.cs — migration's in-epoch capture (HandleMigrate fills the accumulator + inline portion).
• libs/cluster/Server/Migration/MigrateSessionCommonUtils.cs — migration send (per-key Read_UnifiedStore, assemble the captured pieces, WriteOrSend*RecordAsync).
• libs/cluster/Server/Replication/PrimaryOps/DisklessReplication/ReplicationSnapshotIterator.cs — replication send (StoreSnapshotIterator.Reader / WriteRecord / fan-out).
• libs/storage/Tsavorite/cs/src/core/Allocator/ObjectSerialization/ChunkedRecordConstants.cs — the continuation flag.

1. Overview

Both migration and replication ship store records over the network as serialized DiskLogRecord images in a send buffer that is flushed to the target, which deserializes each record back into the store. The two mechanisms differ in how records are produced and applied (see §5): migration retrieves the specific keys being moved (a read per key) and the target applies them conditionally as commands; replication streams the whole store through a snapshot iterator and the replica inserts every record.

• A record whose serialized size fits within NetworkBufferSettings.maxSendBufferContentSize is written whole, as a single non-chunked LogRecord entry (§3).
• A larger record is split into a sequence of chunk entries (ChunkedLogRecord) that the receiver reassembles before deserializing (§4). Because the chunks are reassembled as a ReadOnlySequence<byte> (not one contiguous array), an object value may exceed 2 GB.

This is a serialized DiskLogRecord record image, distinct from the AOF's operation image. For the AOF format see the companion doc, AOF / TsavoriteLog record layout. The two paths share only the chunk continuation flag (ChunkedRecordConstants.ContinuationFlag).

2. Send-buffer framing

The migrate/replica iteration buffer is a single RESP bulk string whose payload is a count-prefixed run of entries (GarnetClientSession.SendAndResetIterationBuffer):

+------------------+----------------------+-----------------------------+------+
| $<len>\r\n       | recordCount (i32)    | entry, entry, entry, ...    | \r\n |
+------------------+----------------------+-----------------------------+------+

Each entry begins with a one-byte MigrationRecordSpanType tag:

This doc covers LogRecord (1) and ChunkedLogRecord (5). Vector-set and RangeIndex encodings are handled separately.

3. Non-chunked record: LogRecord (tag 1)

GarnetClientSession.TryWriteRecordSpan writes the whole record:

+--------+---------------------------------------------+
| type=1 |  serialized DiskLogRecord (§3.1)            |
+--------+---------------------------------------------+

3.1 Serialized DiskLogRecord layout

The whole record is one contiguous image: an inline portion followed by any overflow key, then any overflow value or object value (DiskLogRecord.Serialize / Deserialize). Each overflow key/value is preceded by its full length as a 4-byte little-endian prefix so Deserialize can locate it (the same contiguous layout the chunked form uses). An object value is the tail of the image (no length prefix) and its length is derived on read as recordSpan.Length - objectValueStart, so a small object deserializes directly from the buffer. The inline portion's RDH KeyLength/ValueLength raw fields are left exactly as the source record had them and are never rewritten on the wire. A record is sent whole only when it fits one send buffer; a larger record (including a large object) takes the chunked path (§4).

+=====================================================================+
| INLINE PORTION                                                      |
+----------------+----------------+----------------+------------------+
| RecordInfo     | RecordDataHdr  | ExtendedNamesp.| inline Key       |
| (8 B)          | (8 B, §3.2)    | (0..N B)       | (0..N B)*        |
+----------------+----------------+----------------+------------------+
| inline Value   | optionals: ETag (8 B if present), Expiration       |
| (0..N B)*      | (8 B if present), ObjectLogPosition, ...           |
+----------------+----------------------------------------------------+
| int keyLen (4 B) + OVERFLOW KEY data    (only if the key is Overflow)|
+---------------------------------------------------------------------+
| int valueLen (4 B) + OVERFLOW VALUE data (only if value is Overflow) |
|   -- or -- OBJECT VALUE bytes (no length prefix, only if Object)     |
+=====================================================================+

* For an overflow key/value or object value, the inline slot holds a 4-byte objectId placeholder
  (restored on read); the actual bytes live in the overflow-key / overflow-value / object tail. An
  overflow key/value is preceded by its 4-byte little-endian length; an object value has no prefix.

FixedHeaderSize = RecordInfo.Size (8) + RecordDataHeader.Size (8) = 16. The inline size is FixedHeaderSize + ExtendedNamespaceLength + KeyLength + ValueLength + OptionalSize, rounded up to kRecordAlignment.

3.2 RecordDataHeader (indicator word) — 8 B

A single 8-byte word (atomically published) that fully defines the record's layout (RecordDataHeader.cs):

bit(s)   field
 0       KeyIsInline
 1       ValueIsInline
 2       ValueIsObject
 3       HasExpiration
 4       HasETag
 5       (reserved)
 6..13   FillerWords          (count of 8-byte filler words after alignment padding)
 14..23  KeyLength            (raw inline key length; = objectId size for an overflow key; never rewritten on the wire)
 24..47  ValueLength          (raw inline value length; = objectId size for overflow/object; never rewritten on the wire)
 48..55  RecordType           (byte, caller-interpreted)
 56..63  Namespace            (byte; encodes whether extra namespace bytes precede the key data)

The receiver reads these bits to know whether the key/value are inline, overflow, or (for the value) an object, and therefore how to interpret the tail.

4. Chunked record: ChunkedLogRecord (tag 5)

A record too large for one send buffer — a large inline record, an overflow key/value, or a large object value — is streamed as a sequence of chunk entries. Replication drives this through DiskLogRecord.Serialize (a reused network-mode ChunkedObjectSerializer); migration assembles the captured pieces and sends them via MigrateSession.WriteOrSendAccumulatedRecordAsync. Either way each chunk is framed by GarnetClientSession.TryWriteChunkedRecordSpan:

each chunk entry:
+--------+------------------------------+---------------------+
| type=5 | i32 chunkLength | cont-flag  |  chunk bytes        |
+--------+------------------------------+---------------------+
          bit 31 = ChunkedRecordConstants.ContinuationFlag
                   (1 = more chunks of this record follow; clear on the last chunk)
          bits 30..0 = this chunk's data length

The chunk boundaries are arbitrary send-buffer cut points (not record-structure boundaries), and the chunks of one record may span multiple send buffers / commands. The concatenation of all chunk bytes of a record is the serialized stream:

+=====================+===========+==============+=============+================+==================================+
| INLINE PORTION      | int keyLen| overflow key | int valueLen| overflow value | -- or -- object value (streamed) |
| (RDH, §3.2)         | (4 B)     | (keyLen B)   | (4 B)       | (valueLen B)   |    (no length prefix)            |
+=====================+===========+==============+=============+================+==================================+
   always present        only if key is Overflow      only if value is Overflow    only if value is Object

• The inline portion is the same §3.1 inline bytes (RDH indicator bits tell the receiver which out-of-line components follow).
• Each overflow key/value is preceded by its full length as a 4-byte little-endian prefix, so the receiver can allocate the overflow buffer up front and populate it directly. Overflow key/value are byte[]-bounded (≤ 512 MB), so an int prefix suffices.
• An object value has no length prefix — it is streamed and its length is derived on receive (it may exceed 2 GB, the max length of one byte[]).

4.1 Object value length

An object value's serialized length is not known when the inline portion is emitted, so no length is written for it (on the wire or in the RDH). The receiver derives it as the sum of the accumulated object-value chunk lengths (ChunkedRecordReassembler.objectValueLength) and streams those chunks to the object deserializer; no length is written back into the record's RDH.

4.2 Reassembly and >2 GB support

ChunkedRecordReassembler routes the incoming chunk bytes by component using a small state machine. The inline size comes from DiskLogRecord.GetChunkedRecordInlineSize; the component kinds are read on demand from the record's data header (the RecordDataHeader at the start of the accumulated inline buffer):

• Inline portion: accumulated contiguously into ChunkedRecordReassembler.inlineBuffer; once the fixed header (FixedHeaderSize = 16 B) is present, the layout gives the inline size and which components follow.
• Overflow key / value: the 4-byte length prefix is read, a single OverflowByteArray is allocated up front, and each incoming chunk is copied straight into it (no intermediate list) — populating ChunkedRecordReassembler.keyOverflow / valueOverflow.
• Object value: accumulated as a chunk list (ChunkedRecordReassembler.objectValueChunks, List<byte[]>) and exposed as a ReadOnlySequence<byte> for streaming deserialize — so it may exceed 2 GB.
• Fully-inline record: the whole record is the contiguous inline buffer.

On completion:

• Fully-inline record → DiskLogRecord.Deserialize(inlineBuffer).
• Any out-of-line component → deserialize the object value (if any) from its sequence, then DiskLogRecord.CompleteDeserializeChunkedRecord(header, keyOverflow, valueOverflow, valueObject), which assigns the pre-populated overflow key/value directly (no re-allocation or copy). The RDH length fields are left untouched (the out-of-line lengths rode in the 4-byte wire prefixes, which the receiver already consumed).

5. Send and receive flows

Both mechanisms use the record layout above (a whole LogRecord, or a stream of ChunkedLogRecord chunks), but they produce and apply records differently: migration retrieves the individual keys being moved and applies them conditionally as commands; replication streams the whole store through a snapshot iterator and inserts every record.

5.1 Migration

The source retrieves each key being migrated and captures its pieces in-epoch (HandleMigrate → a MigrationChunkWriterAccumulator): the inline portion is copied into output.SpanByteAndMemory, while the overflow key (a shallow reference — store keys are immutable), the overflow value (a deep copy — the store value may be mutated once the epoch is released), or an object value (serialized into a chunk list, so it may exceed 2 GB) go into the accumulator. Out of epoch the caller assembles [inline][int keyLen][overflow key][int valueLen][overflow value | object chunks] (each overflow key/value preceded by its 4-byte length) and sends it under a CLUSTER MIGRATE command — whole (LogRecord) if the record fits a send buffer, else as ChunkedLogRecord chunks. An object value carries no length prefix; the receiver derives it from the record span (whole records) or the accumulated chunks (chunked), so a small object is sent whole and deserialized directly from the buffer, while a large object chunks. The capture is in-epoch because it runs inside the store's Read callback: Read_UnifiedStore invokes HandleMigrate through ISessionFunctions.Reader, which Tsavorite calls while holding the record's epoch. This in-epoch capture / out-of-epoch send is required because migration sends asynchronously and the store epoch must never be held across an await (unlike replication, which sends synchronously via BlockingWait and can stream a record to the network in-epoch). The target applies each record only if its slot is importing and the key may be written (replace set, or the key is absent).

flowchart TB
  subgraph MSend["Migration send (source)"]
    direction LR
    MA[MIGRATE keys / migrating-slot keys] --> MB[per key: Read_UnifiedStore
capture pieces in-epoch:
inline + accumulator]
    MB --> MC{"larger than send buffer?"}
    MC -->|no| MD[assemble whole
TryWriteRecordSpan type=1 LogRecord]
    MC -->|yes| ME[send pieces as chunks
TryWriteChunkedRecordSpan type=5]
    MD --> MF[(CLUSTER MIGRATE)]
    ME --> MF
  end
  subgraph MRecv["Migration receive (target)"]
    direction LR
    MG[NetworkClusterMigrate:
per record, deserialize] --> MH{"slot importing and replace-or-absent?"}
    MH -->|yes| MI[basicGarnetApi.SET]
    MH -->|no| MJ[skip]
  end
  MF --> MG

Migration also carries VectorSetElement / VectorSetIndex / SerializedRangeIndexStream record kinds (out of scope here); this doc covers only the LogRecord / ChunkedLogRecord kinds.

5.2 Replication

Diskless replica sync takes a streaming-snapshot checkpoint whose StoreSnapshotIterator.Reader pushes every store record to WriteRecord, which fans each record out to all attached replica sessions in lockstep under CLUSTER SYNC. The replica inserts every record unconditionally. A record that fits one send buffer is sent whole (LogRecord): a fully-inline record directly, and a prefix-free object record (inline key + object value) via the chunker's Consume when it drains in one piece (the chunker's ring buffers up to one send buffer, so an object that completes in the first drain fit) — so small objects deserialize directly on the replica. Anything larger streams as ChunkedLogRecord chunks.

flowchart TB
  subgraph RSend["Replication send (primary)"]
    direction LR
    RA[streaming-snapshot checkpoint] --> RB[StoreSnapshotIterator.Reader
pushes each record]
    RB --> RC{"inline and fits buffer?"}
    RC -->|yes| RD[FanOutRecordSpan
type=1 LogRecord]
    RC -->|no| RE[DiskLogRecord.Serialize
drain through chunker]
    RE --> RG2{"fit one drain and prefix-free object?"}
    RG2 -->|yes| RD
    RG2 -->|no| REchunk[FanOutChunk type=5]
    RD --> RF[(CLUSTER SYNC to all replicas)]
    REchunk --> RF
  end
  subgraph RRecv["Replication receive (replica)"]
    direction LR
    RG[NetworkClusterSync:
per record, deserialize] --> RH[basicGarnetApi.SET
every record]
  end
  RF --> RG

5.3 Chunked transfer over the wire (shared)

The chunk framing and reassembly are identical for both mechanisms (only the final apply differs — conditional SET for migration, unconditional SET for replication). A record too large for one send buffer is streamed as ChunkedLogRecord chunks and routed by component in ChunkedRecordReassembler before deserialize:

sequenceDiagram
    participant P as Sender
    participant N as Send buffer
    participant T as Receiver
    participant A as Reassembler
    P->>N: chunk 0, type 5, length with continuation flag, bytes
    P->>N: chunk 1, type 5, length with continuation flag, bytes
    P->>N: chunk M, type 5, length, bytes, continuation flag clear
    N->>T: read entries
    T->>A: Append chunk 0, moreFollow true
    A->>A: route by component: inline, then overflow key/value (into OverflowByteArray), then object chunks
    T->>A: Append chunk M, moreFollow false, complete
    T->>T: CompleteChunkedRecordReassembly: deserialize object value if any
    T->>T: DiskLogRecord Deserialize (fully-inline) or CompleteDeserializeChunkedRecord (assign overflow/object)
    T->>T: SET record into store

5.4 Allocation and copy accounting (reviewer reference)

This subsection enumerates every buffer allocation and byte copy on the record paths and justifies each, for a review focused on minimizing allocations and copies. Three constraints drive the accounting:

1. Migration serializes in-epoch but sends out of epoch (it sends asynchronously, and the store epoch must never be held across an await). Capture is in-epoch because HandleMigrate runs as the store's Read callback (invoked through ISessionFunctions.Reader by Read_UnifiedStore, which Tsavorite calls under epoch protection). Bytes that must outlive the epoch are copied into detached memory before the send; a migrating key is not locked, so its value may change concurrently.
2. Replication sends synchronously in-epoch (it flushes via BlockingWait and never awaits), so it can stream a record straight from the record's native log memory into the send buffer, with no per-record heap copy.
3. An overflow key/value has a known length and lands in a single owned buffer, while an object value has an unknown, possibly larger-than-2 GB length and is held as a list of chunks (List<byte[]>), never one array.

Inline portion - when it is copied. The inline portion (RecordInfo + RDH + inline key/value + optionals, padded to RoundUp(ActualSize)) is copied only under these conditions:

• Migration, always (DirectCopyInlinePortionOfRecord into output.SpanByteAndMemory, in-epoch). Required because the record's native log memory is only valid in-epoch and the send happens after the epoch is released. The emitted bytes are the compacted component image (RoundUp(ActualSize)), so the receiver locates the overflow components at RoundUp(ActualSize) - computed from the RDH component lengths and therefore filler-independent. The output buffer is reused across keys (heap-backed through MemoryPool only when it must outlive the network buffer, e.g. an object value).
• Replication whole-inline fast path (DirectCopyInlinePortionOfRecord into the reused serializationOutput). Copied once, then fanned to each replica's send buffer. Required to stage one stable, compacted image that is copied to all N replica buffers without re-reading the (possibly-evicted) record per replica.
• Replication chunked path - not separately copied. SerializeChunked emits the inline portion directly from the record's native memory (chunker.WriteBytes over physicalAddress); the only staging is the chunker ring, allocated once and reused across all records.

The filler word does not appear above because it is immaterial to the receiver, which never reads it: each record is applied with a semantic Upsert that sizes the destination from the source's key/value field info, not from the wire image (GetUpsertRecordSize: "Filler information is not needed"). Only the RoundUp(ActualSize) byte count matters on the wire, and both send paths emit exactly that. DirectCopyInlinePortionOfRecord does rewrite the copied record's filler word, but not for the wire path: it is a shared helper also used by in-memory RENAME (HandleRename), where the output buffer becomes a live LogRecord that must carry a self-consistent filler; the migration and replication inline paths inherit that rewrite for free. SerializeChunked streams straight from the live record and skips the rewrite (which would otherwise require staging the record in a scratch copy first), safe because the receiver ignores the wire filler.

In short, the inline portion is copied once when the bytes must outlive the producing epoch (migration) or be fanned identically to multiple replicas (replication fast path); it is streamed with no extra allocation when a single synchronous consumer drains it in-epoch (replication chunked path, through the reused ring).

Chunk accumulation - where and why. Chunks are accumulated as a List<byte[]> in exactly two places, both only for an object value:

• Migration send (MigrationChunkWriterAccumulator.objectValueChunks). The object is serialized in-epoch through a reused 4 MB ring; each drain is copied into an owned chunk. Required because (a) the live object may change once the epoch is released, so it must be snapshotted in-epoch, and (b) a serialized object may exceed 2 GB, which a single byte[] cannot hold. The ring is allocated once per migration (reused across keys); only the per-object chunk arrays are per-record. An overflow value on this path is instead a single deep-copy array (SetValueOverflowDeepCopy), and an overflow key is a shallow reference to the store's immutable array (no copy).
• Receive (ChunkedRecordReassembler.objectValueChunks). Each arriving object chunk is copied once from the transient network buffer into an owned array, because the object length is not known up front (no single array can be pre-sized) and may exceed 2 GB. The chunks are wrapped as a ReadOnlySequence<byte> (no further copy) and streamed to the object deserializer.

Everything else avoids accumulation. On receive, an overflow key/value is a single up-front OverflowByteArray (sized from its 4-byte length prefix) that chunk bytes are copied directly into (FillOverflow), never staged in an intermediate buffer; CompleteDeserializeChunkedRecord then assigns it to the record with no re-copy. Replication send never accumulates - it streams synchronously in-epoch through the shared reused ring, so an object value is never materialized whole on the sender.

Copies at a glance.

[1] Assembled whole (rather than chunked) so the receiver reads one type-1 LogRecord entry instead of reassembling multiple chunks; the trade-off is detailed in One extra copy, called out below.

[2] Backpressure here means the client's send buffer filled mid-record, so it must be flushed (drained to the socket) before the chunk can be re-sent. span.ToArray() detaches the chunk first because the reused send span may not survive the flush await.

Every copy that remains after the table above is the unavoidable transfer into or out of the transient network send/receive buffer.

One extra copy, called out. On the migration path only, a non-inline record that fits one send buffer is assembled into one contiguous LogRecord entry, so its already-detached pieces (inline portion, overflow value or object chunks) are copied once more into sendAssembleBuffer before the send. This is the price of emitting a single type-1 entry from non-contiguous captured pieces; the alternative (chunking a buffer-sized record) would trade this copy for a multi-chunk reassembly on the receiver. The extra copy is bounded by the send-buffer size and does not apply to a fully-inline record (sent straight from output.SpanByteAndMemory) or to the replication paths.

6. Call sequence (code paths)

Indentation = call depth; a multi-step flow may sit on one line (a → b → c), and italic sub-items are terse notes. Migration and replication differ on write; they share the receive path. The record's components — inline portion, overflow key, overflow value, object value — are called out throughout.

Migration write — retrieve each key from the store and send it (in-epoch capture → out-of-epoch send):

• MigrateSession.TransmitKeysAsync()
  • MigrateOperation.cs; iterate the keys being migrated
  • WriteOrSendRecordAsync(key)
    • MigrateSessionCommonUtils.cs
    • BasicGarnetApi.Read_UnifiedStore(key) → HandleMigrate(srcLogRecord)
      • UnifiedStore/ReadMethods.cs; capture the record's pieces in-epoch (a migrating key is not locked, so its value may change concurrently)
      • DiskLogRecord.SerializeInlinePortionForMigration()
        • inline portion → output.SpanByteAndMemory; copy only; overflow key/value lengths are added as 4-byte prefixes by the sender out of epoch
      • acc.SetKeyOverflow(KeyOverflow)
        • overflow key → shallow ref (store keys are immutable); populates MigrationChunkWriterAccumulator.keyOverflow
      • acc.SetValueOverflowDeepCopy(ValueOverflow)
        • overflow value → deep copy (the value may change once the epoch is released); populates MigrationChunkWriterAccumulator.valueOverflow
      • acc.SerializeObjectValue(ValueObject) → ChunkedObjectSerializer → acc.Consume(first, second)
        • object value → serialized chunks; populates MigrationChunkWriterAccumulator.objectValueChunks (List<byte[]>, supports values over 2 GB). The serializer is reused across records, so its ring is allocated once per migration, not per object
    • WriteOrSendAccumulatedRecordAsync(inline, acc)
      • out of epoch: assemble the captured pieces and send
      • gcs.TryWriteRecordSpan(record, LogRecord)
        • fits one buffer → whole record (type 1); assemble [inline][int keyLen][overflow key][int valueLen][overflow value | object bytes] contiguously (each overflow key/value preceded by its 4-byte length; an object value is the tail, length derived on read from the record span)
      • gcs.TryWriteChunkedRecordSpan(chunk, moreFollow)
        • too large for one buffer → chunks (type 5): stream [inline][int keyLen][overflow key][int valueLen][overflow value | object chunks], continuation until the last byte → send buffer → CLUSTER MIGRATE

Replication write — stream a store snapshot to the replica(s) (in-epoch, synchronous send):

• ReplicationSyncManager → StoreSnapshotIterator.Reader(srcLogRecord) → SnapshotIteratorManager.WriteRecord(srcLogRecord)
  • ReplicationSnapshotIterator.cs; a streaming-snapshot checkpoint pushes each record
  • DiskLogRecord.DirectCopyInlinePortionOfRecord() → FanOutRecordSpan(LogRecord)
    • inline record → whole (type 1), fanned to all replicas in lockstep
  • DiskLogRecord.Serialize(srcLogRecord, serializer, chunker) → SnapshotIteratorManager.Consume(...)
    • DiskLogRecord.cs; non-inline → stream through the chunker (epoch held; replication flushes synchronously via BlockingWait, so it may stream in-epoch)
    • writes the inline portion, then (each preceded by its 4-byte length) the overflow key and overflow value — or streams the object value (no prefix) — into the chunk stream; the carrier is the streamed netbuffer chunk (no accumulator)
    • Consume: first drain is isComplete and prefix-free object → FanOutRecordSpan(LogRecord)
      • small object (inline key + object value) that fit one drain → whole record (type 1), deserialized directly on the replica
    • Consume: otherwise → FanOutChunk(chunk, moreFollow)
      • too large for one buffer → chunks (type 5), continuation until the last byte → send buffer → CLUSTER SYNC

Receive (migration + replication) — read records from the buffer and write them to the store:

• NetworkClusterMigrate() / NetworkClusterSync()
  • RespClusterMigrateCommands.cs / RespClusterReplicationCommands.cs; per record in the payload
  • type == LogRecord → DiskLogRecord.Deserialize(recordSpan)
    • whole record: inline + any overflow key/value (each preceded by its 4-byte length) or a small object value (length derived from the record span), restored/deserialized directly from the contiguous span
  • type == ChunkedLogRecord → ChunkedRecordReassembler.Append(chunk, moreFollow)
    • route bytes by component (may span commands): inline → inlineBuffer; overflow key/value → a single OverflowByteArray (keyOverflow/valueOverflow) populated directly from the chunks; object value → objectValueChunks (List<byte[]>)
    • CompleteChunkedRecordReassembly(headerPtr)
      • component kinds read from the record header in inlineBuffer; a fully-inline record → DiskLogRecord.Deserialize(inlineBuffer)
      • GarnetObjectSerializer.Deserialize(ObjectValueSequence()) → DiskLogRecord.CompleteDeserializeChunkedRecord(header, keyOverflow, valueOverflow, valueObject)
        • assign the pre-populated overflow key/value directly + the deserialized (over-2 GB-capable) object value; the RDH lengths are left untouched (the 4-byte wire prefixes were already consumed)
  • basicGarnetApi.SET(diskLogRecord)
    • → store (migrate: only if the slot is importing and replace-or-absent; replication: unconditional)