Krypton Hybrid

Krypton Hybrid

Overview

Krypton Hybrid is a fork based on Krypton fnp legacy（1.19.2),Completed the missing optimizations in the Krypton fnp lite version, added support for the zstd compression algorithm, and made improvements to the original encoding and Netty program processing.while retaining the features of Krypton fnp and RecastLib.

Differentiation

1. Compatibility support for hybrid servers is offered.

2. Commands are provided to monitor traffic usage.

3. This project prioritizes using zstd as the primary compression algorithm, while also providing optimizations for entity tracking and communication encoding from the original version.

Technical Details

1. Network Compression Pipeline

Krypton Hybrid replaces Minecraft's default zlib codec by injecting into Connection.setupCompression.

Feature	Detail
Primary algorithm	Zstd via zstd-jni (native)
Fallback algorithm	zlib via velocity-native
Level range	1 – 22 (default 3)
Multi-threaded	workers ≥ 1 activates native thread pool
Dictionary	Optional .zdict pre-trained file, fail-open or fail-fast
Advanced tuning	overlap_log, job_size, strategy, LDM + window_log

Both server and client must use the same algorithm. A mismatch corrupts the session immediately.

---

2. Light Payload Optimization — Uniform-RLE

Sky-light sections above the surface are almost always uniform (0xFF = all-15). Instead of writing 2 048 bytes per uniform layer, Krypton encodes them as 2 bytes.

Wire format (ClientboundLightUpdatePacketData):

[0x4B]            – Krypton marker (1 byte)
[skyYMask]        – BitSet
[blockYMask]      – BitSet
[emptySkyYMask]   – BitSet
[emptyBlockYMask] – BitSet
[skyCount]        – VarInt
For each sky DataLayer:
  [0x01] + 1 byte     – uniform (all nibbles == byte)
  [0x00] + 2048 bytes – raw    (fixed size, no VarInt prefix)
[blockCount]      – VarInt
For each block DataLayer: same as above

Savings guard: only activates when 2048 × uniform_count + raw_count − 1 > 0. Max saving: ~40 KB per chunk in open-sky environments (view-distance 12 → up to 20 uniform sky layers).

---

3. Chunk Data Optimization — XOR-Delta Heightmaps + Biome Stream

Wire format (ClientboundLevelChunkPacketData):

[0x4B]            – Krypton chunk-data marker (1 byte)

──── Heightmaps (binary + XOR-delta) ────
VarInt  entry_count
Per entry:
  UTF-8  key
  VarInt long_count
  long[] XOR-delta encoded  (delta[i] = raw[i] ^ raw[i-1])

──── Block data (sections without biomes) ────
VarInt  blocks_length
byte[]  blocks  (short nonEmptyCount + PalettedContainer<BlockState> per section)

──── Biome data (compact stream) ────
VarInt  section_count
Per section:
  0x01 + VarInt(biomeId)          – single-value section (2 bytes vs. 3+)
  0x00 + VarInt(len) + raw bytes  – multi-value section

──── Block entities ────
LIST_STREAM_CODEC (unchanged vanilla encoding)

Optimization	Mechanism	Benefit
XOR-delta heightmaps	Adjacent columns correlate → many near-zero longs	Better Zstd/zlib ratio + ~40 B NBT overhead eliminated
Biome stream split	Biome data grouped; single-value sections compacted	Single-value: 2 B vs. 3+ B; cross-section redundancy exploited

---

4. Delayed Chunk Cache (DCC)

Departing chunks are buffered instead of immediately dropped. If the player re-enters range before the entry expires, the full resend is skipped.

Config key	Default	Description
dcc.enabled	true	Master switch
dcc.size_limit	60	Max buffered chunks per player
dcc.distance	5	Cache radius around player position (chunks)
dcc.timeout_seconds	30	Forced eviction timeout

Eviction occurs on three dimensions: distance, capacity, and timeout. Evicted entries trigger deferred dropChunk callbacks on the next ChunkMap.tick().

---

5. Flush Consolidation + Entity Packet Bundling

Two complementary layers reduce syscall and framing overhead during the entity tracking tick:

ChunkMap.tick() HEAD
  ① Recovery: flush any stale batch from previous failed tick
  ② Disable auto-flush for all players
  ③ Open EntityBundleCollector batch window

  ↳ TrackedEntity.broadcast() → packets collected per player

ChunkMap.tick() RETURN
  ④ Close batch → emit ClientboundBundlePackets (buffered, not yet flushed)
  ⑤ Re-enable auto-flush → single kernel flush per player

Layer	Mechanism	Benefit
Flush consolidation (Netty)	Buffer writes, one flush() at end of tick	Fewer syscalls on busy servers
Bundle coalescing (protocol)	N entity packets → 1 ClientboundBundlePacket	Reduced framing overhead; better compression ratio; client-side atomicity

---

6. Hot-Path Micro-Optimizations

Target	Optimization
VarInt.read()	Branchless 4-byte getIntLE + bit-twiddling fast path (covers values 0 – 268 M)
VarInt.getByteSize()	Lookup-table via Integer.numberOfLeadingZeros
VarLong.getByteSize()	Same lookup-table approach
FriendlyByteBuf.writeVarInt()	Peeled 1/2/3/4/5-byte paths, writes directly to backing ByteBuf
FriendlyByteBuf.writeUtf()	Single-pass via ByteBufUtil.utf8Bytes — no intermediate byte[]
Varint21FrameDecoder	Velocity-derived efficient framing + nullping attack mitigation

---

7. Observability — /krypton Commands

/krypton stats show          – compression ratio, bandwidth saving, uptime
/krypton stats reset         – reset all counters
/krypton packets bycount [n] – top N packet types by count
/krypton packets bybytes [n] – top N packet types by bytes
/krypton mods bycount [n]    – top N mod namespaces by count
/krypton mods bybytes [n]    – top N mod namespaces by bytes

Requires operator permission level 2.

test data

In the original test, this mod was able to reduce the initial package size to 13% of the original.

Credit

• Krypton Fabric
• Velocity
• VelocityNT Recast
• Paper
• RecastSSL
• NotEnoughBandwidth