Chunk Format

This article describes in additional detail the format of the Chunk Data packet.

Concepts

Chunks columns and Chunk sections

You've probably heard the term "chunk" before. Minecraft uses chunks to store and transfer world data. However, there are actually 2 different concepts that are both called "chunks" in different contexts: chunk columns and chunk sections.

A chunk column is a 16×256×16 collection of blocks, and is what most players think of when they hear the term "chunk". However, these are not the smallest unit data is stored in in the game; chunk columns are actually 16 chunk sections aligned vertically.

Chunk columns store biomes, block entities, entities, tick data, and an array of sections.

A chunk section is a 16×16×16 collection of blocks (chunk sections are cubic). This is the actual area that blocks are stored in, and is often the concept Mojang refers to via "chunk". Breaking columns into sections wouldn't be useful, except that you don't need to send all chunk sections in a column: If a section is empty, then it doesn't need to be sent (more on this later).

Chunk sections store blocks and light data (both block light and sky light). Additionally, they can be associated with a section palette. A chunk section can contain at maximum 4096 (16×16×16, or 2¹²) unique IDs (but, it is highly unlikely that such a section will occur in normal circumstances).

Chunk columns and chunk sections are both displayed when chunk border rendering is enabled (F3+G). Chunk columns borders are indicated via the red vertical lines, while chunk sections borders are indicated by the blue lines.

Empty sections and the primary bit mask

As previously mentioned, chunk sections can be empty. Sections which contain no useful data are treated as empty, and are not sent to the client, as the client is able to infer the contents. For the average world, this means around 60% of the world's data doesn't need to be sent, since it's all air; this is a significant save.

It is important to note that a chunk composed entirely of empty sections is different from an empty (ie, unloaded) chunk column. When a block is changed in an empty section, the section is created (as all air), and the block is set. When a block is changed in an empty chunk, the behavior is undefined (but generally, nothing happens).

The primary bit mask simply determines which sections are being sent. The least significant bit is for the lowest section (y=0 to y=15). Only 16 bits can be set in it (with the 16th bit controlling the y=240 to y=255 section); sections above y=255 are not valid for the notchian client. To check whether a section is included, use ((mask & (1 << sectionY)) != 0).

Global and section palettes

Illustration of an indexed palette (Source)

Minecraft also uses palettes. A palette maps numeric IDs to block states. The concept is more commonly used with colors in an image; Wikipedia's articles on color look-up tables, indexed colors, and palettes in general may be helpful for fully grokking it.

There are 2 palettes that are used in the game: the global palette and the section palette.

The global palette is the standard mapping of IDs to block states. Currently, it is a combination Block ID and Metadata ((blockId << 4) | metadata). Note that thus, the global palette is not continuous. Entries not defined within the global palette are treated as air (even if the block ID itself is known, if the metadata is not known, the state is treated as air). Note that the global palette is currently represented by 13 bits per entry, with 9 bits for the block ID and 4 bits for the metadata.

The basic implementation looks like this:

long getGlobalPaletteIDFromState(BlockState state) {
    if (state.isValid()) {
        return (state.getId() << 4) | state.getMetadata();
    } else {
        return 0;
    }
}

BlockState getStateFromGlobalPaletteID(long id) {
    int blockID = (id >> 4);
    byte metadata = (id & 0x0F);
    BlockState state = new BlockState(blockID, metadata);
    if (state.isValid()) {
        return state;
    } else {
        return new BlockState(0, 0);  // Air
    }
}

Don't assume that the global palette will always be like this; keep it in a separate function. Mojang has stated that they plan to change the global palette to avoid increasing the total size. Equally so, though, do not hardcode the total size of the palette; keep it in a constant.

A section palette is used to map IDs within a chunk section to global palette IDs. Other than skipping empty sections, correct use of the section palette is the biggest place where data can be saved. Given that most sections contain only a few blocks, using 13 bits to represent a chunk section that is only stone, gravel, and air would be extremely wasteful. Instead, a list of IDs are sent mapping indexes to global palette IDs (for instance, 0x10 0xD0 0x00), and indexes within the section palette are used (so stone would be sent as 0, gravel 1, and air 2). The number of bits per ID in the section palette varies from 4 to 8; if fewer than 4 bits would be needed it's increased to 4 and if more than 8 would be needed, the section palette is not used and instead global palette IDs are used.

Note that the notchian client (and server) store their chunk data within the compacted, paletted format. Sending non-compacted data not only wastes bandwidth, but also leads to increased memory use clientside; while this is OK for an initial implementation it is strongly encouraged that one compacts the block data as soon as possible.

Ground-up continuous

The ground-up continuous value (tentative name) is one of the more confusing properties of the chunk data packet, simply because there's no good name for it. It controls two different behaviors of the chunk data packet, one that most people need, and one that most don't.

When ground-up continuous is set, the chunk data packet is used to create a new chunk. This includes biome data, and all (non-empty) sections in the chunk. Sections not specified in the primary bit mask are empty sections.

Sending a packet with ground-up continuous enabled over a chunk that already exists will leak memory clientside.

Make sure to unload chunks before overwriting them with the Unload Chunk packet. That packet can always be sent even on unloaded chunks, so in situations where the chunk might or might not be loaded already, it's valid to send it again (but avoid sending it in excess).

The MultiplayerChunkCache values in F3 show the number of chunks in the client's 2 chunk storage mechanisms; if the numbers aren't equal, you've leaked chunks.

When ground-up continuous is not set, then the chunk data packet acts as a large Multi Block Change packet, changing all of the blocks in the given section at once. This can have some performance benefits, especially for lighting purposes. BIome data is not sent when ground-up continuous is not set; that means that biomes can't be changed once a chunk is loaded. Sections not specified in the primary bit mask are not changed and should be left as-is.

As with Multi Block Change and Block Change, it is not safe to send this packet in unloaded chunks, as it can corrupt notchian client's shared empty chunk. Clients should ignore such packets, and servers should not send non-ground-up continuous chunk data packets into unloaded chunks.

Packet structure

Packet ID	State	Bound To	Field Name	Field Type	Notes
0x20	Play	Client	Chunk X	Int	Chunk coordinate (block coordinate divided by 16, rounded down)
			Chunk Z	Int	Chunk coordinate (block coordinate divided by 16, rounded down)
			Ground-Up Continuous	Boolean	See #Ground-up continuous
			Primary Bit Mask	VarInt	Bitmask with bits set to 1 for every 16×16×16 chunk section whose data is included in Data. The least significant bit represents the chunk section at the bottom of the chunk column (from y=0 to y=15).
			Size	VarInt	Size of Data in bytes
			Data	Byte array	See data structure below
			Number of block entities	VarInt	Length of the following array
			Block entities	Array of NBT Tag	All block entities in the chunk. Use the x, y, and z tags in the NBT to determine their positions.

Data structure

The data section of the packet contains most of the useful data for the chunk.

Field Name	Field Type	Notes
Data	Array of Chunk Section	The length of the array is equal to the number of bits set in Primary Bit Mask. Sections are sent bottom-to-top, i.e. the first section, if sent, extends from Y=0 to Y=15.
Biomes	Optional Byte Array	Only sent if Ground-Up Continuous is true; 256 bytes if present

Chunk Section structure

A Chunk Section is defined in terms of other data types. A Chunk Section consists of the following fields:

Field Name	Field Type	Notes
Bits Per Block	Unsigned Byte	Determines how many bits are used to encode a block. Note that not all numbers are valid here. This also changes whether the palette is present.
Palette Length	VarInt	Length of the following array. May be 0, in which case the following palette is not sent.
Palette	Optional Array of VarInt	Mapping of block state IDs in the global palette to indices of this array
Data Array Length	VarInt	Number of longs in the following array
Data Array	Array of Long	Compacted list of 4096 indices pointing to state IDs in the Palette
Block Light	Byte Array	Half byte per block
Sky Light	Optional Byte Array	Only if in the Overworld; half byte per block

Data Array, Block Light, and Sky Light are given for each block with increasing x coordinates, within rows of increasing z coordinates, within layers of increasing y coordinates. For the half-byte light arrays, even-indexed items (those with an even x coordinate, starting at 0) are packed into the low bits, odd-indexed into the high bits.

The Data Array, although varying in length, will never be padded due to the number of blocks being evenly divisible by 64, which is the number of bits in a long.

There are several values that can be used for the bits per block value. In most cases, invalid values will be interpreted as a different value when parsed by the notchian client, meaning that chunk data will be parsed incorrectly if you use an invalid bits per block. Servers must make sure that the bits per block value is correct.

• up to 4: Blocks are encoded as 4 bits. The palette is used and sent.
• 5 to 8: Blocks are encoded with the given number of bits. The palette is used and sent.
• 9 and above: The palette is not sent. Blocks are encoded by their whole ID in the global palette, with bits per block being set as the base 2 logarithm of the number of block states, rounded up. For the current vanilla release, this is 13 bits per block.

The global palette encodes a block as 13 bits. It uses the block ID for the first 9 bits, and the block damage value for the last 4 bits. For example, Diorite (block ID 1 for minecraft:stone with damage 3) would be encoded as 000000001 0011. If a block is not found in the global palette (either due to not having a valid damage value or due to not being a valid ID), it will be treated as air.

If Minecraft Forge is installed and a sufficiently large number of blocks are added, the bits per block value for the global palette will be increased to compensate for the increased ID count. This increase can go up to 16 bits per block (for a total of 4096 block IDs; when combined with the 16 damage values, there are 65536 total states). You can get the number of blocks with the "Number of ids" field found in the RegistryData packet in the Forge Handshake.

The data array stores several entries within a single long, and sometimes overlaps one entry between multiple longs. For a bits per block value of 13, the data is stored such that bits 1 through 13 are the first entry, 14 through 26 are the second, and so on. Note that bit 1 is the least significant bit in this case, not the most significant bit. The same behavior applies when a value stretches between two longs: for instance, block 5 would be bits 53 through 64 of the first long and then bit 65 of the second long.

Example

13 bits per block, using the global palette.

The following two longs would represent...

1001880C0060020 = 0000000100000000000110001000000011000000000001100000000000100000
200D0068004C020 = 0000001000000000110100000000011010000000000001001100000000100000

9 blocks, with the start of a 10th (that would be finished in the next long).

1. Grass, 2:0
2. Dirt, 3:0
3. Dirt, 3:0
4. Coarse dirt, 3:1
5. Stone, 1:0
6. Stone, 1:0
7. Diorite, 1:3
8. Gravel, 13:0
9. Gravel, 13:0
10. Stone, 1:0 (or potentially emerald ore, 129:0)

Biomes

The biomes array is only present when ground-up continuous is set to true. Biomes cannot be changed unless a chunk is re-sent.

The structure is an array of 256 bytes, each representing a Biome ID (it is recommended that 127 for "Void" is used if there is no set biome). The array is indexed by z * 16 | x.

Tips

There are several things that can make it easier to implement this format.

• The 13 value for full bits per block is likely to change in the future, so it should not be hardcoded (instead, it should either be calculated or left as a constant).
• Servers do not need to implement the palette initially (instead always using 13 bits per block), although it is an important optimization later on.
• The Notchian server implementation does not send values that are out of bounds for the palette. If such a value is received, the format is being parsed incorrectly. In particular, if you're reading a number with all bits set (15, 31, etc), then you're probably reading sky light data.
• The number of longs needed for the data array can be calculated as ((16×16×16 blocks)×Bits per block)÷64 bits per long (which simplifies to 64×Bits per block). For instance, 13 bits per block requires 832 longs.

Sample implementations

How the chunk format can be implemented varies largely by how you want to read/write it. It is often easier to read/write the data long-by-long instead of pre-create the data to write; however, storing the chunk data arrays in their packed form can be far more efficient memory- and performance-wise. These implementations are simple versions that can work as a base (especially for dealing with the bit shifting), but are not ideal.

Deserializing

When deserializing, it is easy to read to a buffer (since length information is present). A basic example:

public Chunk ReadChunkDataPacket(Buffer data) {
    int x = ReadInt(data);
    int z = ReadInt(data);
    bool full = ReadBool(data);
    Chunk chunk;
    if (full) {
        chunk = new Chunk(x, z);
    } else {
        chunk = GetExistingChunk(x, z);
    }
    int mask = ReadVarInt(data);
    int size = ReadVarInt(data);
    ReadChunkColumn(chunk, full, mask, data.ReadByteArray(size));

    int blockEntityCount = ReadVarInt(data);
    for (int i = 0; i < blockEntityCount; i++) {
        CompoundTag tag = ReadCompoundTag(data);
        chunk.AddBlockEntity(tag.GetInt("x"), tag.GetInt("y"), tag.GetInt("z"), tag);
    }

    return chunk;
}

private void ReadChunkColumn(Chunk chunk, bool full, int mask, Buffer data) {
    for (int sectionY = 0; sectionY < CHUNK_HEIGHT / SECTION_HEIGHT; y++) {
        if ((mask & (1 << sectionY)) != 0) {  // Is the given bit set in the mask?
            byte bitsPerBlock = ReadByte(data);

            // Excessively specific format that exactly matches the client logic
            // This extra checking makes sense on the server side, but client
            // side it only is needed when dealing with servers sending incorrect packets
            // (the notchian server will not send such packets)
            if (bitsPerBlock < 4) {
                bitsPerBlock = 4;
            }
            if (bitsPerBlock > 8) {
                bitsPerBlock = FULL_SIZE_BITS_PER_BLOCK;  // 13, currently, but liable to eventually change
            }

            bool usePalette = (bitsPerBlock <= 8)

            int[] palette = null;
            if (usePalette) {
                int numPaletteEntries = ReadVarInt(data);
                palette = new int[numPaletteEntries];
                for (int i = 0; i < numPaletteEntries; i++) {
                    palette[i] = ReadVarInt(data);
                }
            } else {
                ReadVarInt(data);  // Should always be 0
            }

            // A bitmask that contains bitsPerBlock set bits
            uint individualValueMask = (uint)((1 << bitsPerBlock) - 1);

            int dataArrayLength = ReadVarInt(data);
            UInt64[] dataArray = ReadUInt64Array(data, dataArrayLength);

            ChunkSection section = new ChunkSection();

            for (int y = 0; y < SECTION_HEIGHT; y++) {
                for (int z = 0; z < SECTION_WIDTH; z++) {
                    for (int x = 0; x < SECTION_WIDTH; x++) {
                        int blockNumber = (((blockY * SECTION_HEIGHT) + blockZ) * SECTION_WIDTH) + blockX;
                        int startLong = (blockNumber * bitsPerBlock) / 64;
                        int startOffset = (blockNumber * bitsPerBlock) % 64;
                        int endLong = ((blockNumber + 1) * bitsPerBlock - 1) / 64;

                        uint data;
                        if (startLong == endLong) {
                            data = (uint)(dataArray[startLong] >> startOffset);
                        } else {
                            int endOffset = 64 - startOffset;
                            blockId = (uint)(dataArray[startLong] >> startOffset | dataArray[endLong] << endOffset);
                        }
                        data &= individualValueMask;

                        if (usePalette) {
                            // data should always be within the palette length
                            // If you're reading a power of 2 minus one (15, 31, 63, 127, etc...) that's out of bounds,
                            // you're probably reading light data instead
                            data = palette[data];
                        }

                        BlockState state = GetStateFromGlobalPaletteID(data);
                        section.SetState(x, y, z, state);
                    }
                }
            }

            for (int y = 0; y < SECTION_HEIGHT; y++) {
                for (int z = 0; z < SECTION_WIDTH; z++) {
                    for (int x = 0; x < SECTION_WIDTH; x += 2) {
                        // Note: x += 2 above; we read 2 values along x each time
                        byte value = ReadByte(data);

                        section.SetBlockLight(x, y, z, value & 0xF);
                        section.SetBlockLight(x + 1, y, z, (value >> 4) & 0xF);
                    }
                }
            }

            if (currentDimension.HasSkylight()) { // IE, current dimension is overworld / 0
                for (int y = 0; y < SECTION_HEIGHT; y++) {
                    for (int z = 0; z < SECTION_WIDTH; z++) {
                        for (int x = 0; x < SECTION_WIDTH; x += 2) {
                            // Note: x += 2 above; we read 2 values along x each time
                            byte value = ReadByte(data);

                            section.SetSkyLight(x, y, z, value & 0xF);
                            section.SetSkyLight(x + 1, y, z, (value >> 4) & 0xF);
                        }
                    }
                }
            }

            // May replace an existing section or a null one
            chunk.Sections[SectionY] = section;
        }
    }

    for (int z = 0; z < SECTION_WIDTH; z++) {
        for (int x = 0; x < SECTION_WIDTH; x++) {
            chunk.SetBiome(x, z, ReadByte(data));
        }
    }
}

// Value should already have gone through the section palette
private BlockState GetStateFromGlobalPaletteID(uint value) {
    // This method is subject to change in future MC versions

    byte metadata = data & 0xF;
    uint id = data >> 4;

    return BlockState.ForIDAndMeta(id, metadata);
}

Serializing

Serializing the packet is more complicated, because of the palette. It is easy to implement with the full bits per block value; implementing it with a compacting palette is much harder since algorithms to generate and resize the palette must be written. As such, this example does not generate a palette. The palette is a good performance improvement (as it can significantly reduce the amount of data sent), but managing that is much harder and there are a variety of ways of implementing it.

Also note that this implementation doesn't handle situations where full is false (ie, making a large change to one section); it's only good for serializing a full chunk.

public void WriteChunkDataPacket(Chunk chunk, Buffer data) {
    WriteInt(data, chunk.GetX());
    WriteInt(data, chunk.GetZ());
    WriteBool(true);  // Full

    int mask = 0;
    Buffer columnBuffer = new Buffer();
    for (int sectionY = 0; sectionY < CHUNK_HEIGHT / SECTION_HEIGHT; y++) {
        if (!chunk.IsSectionEmpty(sectionY)) {
            mask |= (1 << chunkY);  // Set that bit to true in the mask
            WriteChunkSection(chunk.Sections[sectionY], columnBuffer);
        }
    }
    for (int z = 0; z < SECTION_WIDTH; z++) {
        for (int x = 0; x < SECTION_WIDTH; x++) {
            WriteByte(columnBuffer, chunk.GetBiome(x, z));  // Use 127 for 'void' if your server doesn't support biomes
        }
    }

    WriteVarInt(data, mask);
    WriteVarInt(data, columnBuffer.Size);
    WriteByteArray(data, columnBuffer);

    // If you don't support block entities yet, use 0
    // If you need to implement it by sending block entities later with the update block entity packet,
    // do it that way and send 0 as well.  (Note that 1.10.1 (not 1.10 or 1.10.2) will not accept that)

    WriteVarInt(data, chunk.BlockEntities.Length);
    foreach (CompoundTag tag in chunk.BlockEntities) {
        WriteCompoundTag(data, tag);
    }
}

private void WriteChunkSection(ChunkSection section, Buffer data) {
    byte bitsPerBlock = FULL_SIZE_BITS_PER_BLOCK;  // 13

    WriteByte(bitsPerBlock);
    WriteVarInt(data, 0);  // Palette size is 0

    int dataLength = (16*16*16) * bitsPerBlock / 64; // See tips section for an explanation of this calculation
    UInt64[] data = new UInt64[dataLength];

    // A bitmask that contains bitsPerBlock set bits
    uint individualValueMask = (uint)((1 << bitsPerBlock) - 1);

    for (int y = 0; y < SECTION_HEIGHT; y++) {
        for (int z = 0; z < SECTION_WIDTH; z++) {
            for (int x = 0; x < SECTION_WIDTH; x++) {
                int blockNumber = (((blockY * SECTION_HEIGHT) + blockZ) * SECTION_WIDTH) + blockX;
                int startLong = (blockNumber * bitsPerBlock) / 64;
                int startOffset = (blockNumber * bitsPerBlock) % 64;
                int endLong = ((blockNumber + 1) * bitsPerBlock - 1) / 64;

                BlockState state = section.GetState(x, y, z);

                uint value = GetGlobalPaletteIDFromState(state);
                value &= individualValueMask;

                data[startLong] |= (Value << startOffset);

                if (startLong != endLong) {
                    data[endLong] = (value >> (64 - startOffset));
                }
            }
        }
    }

    WriteVarInt(dataLength);
    WriteLongArray(data);

    for (int y = 0; y < SECTION_HEIGHT; y++) {
        for (int z = 0; z < SECTION_WIDTH; z++) {
            for (int x = 0; x < SECTION_WIDTH; x += 2) {
                // Note: x += 2 above; we read 2 values along x each time
                byte value = section.GetBlockLight(x, y, z) | (section.GetBlockLight(x + 1, y, z) << 4);
                WriteByte(data, value);
            }
        }
    }

    if (currentDimension.HasSkylight()) { // IE, current dimension is overworld / 0
        for (int y = 0; y < SECTION_HEIGHT; y++) {
            for (int z = 0; z < SECTION_WIDTH; z++) {
                for (int x = 0; x < SECTION_WIDTH; x += 2) {
                    // Note: x += 2 above; we read 2 values along x each time
                    byte value = section.GetSkyLight(x, y, z) | (section.GetSkyLight(x + 1, y, z) << 4);
                    WriteByte(data, value);
                }
            }
        }
    }
}

private uint GetGlobalPaletteIDFromState(BlockState state) {
    // NOTE: This method will probably change in new versions
    byte metadata = state.getMetadata();
    uint id = section.GetBlockID(x, y, z);

    return id << 4 | metadata;
}

Full implementations

• Java, 1.12.2, writing only, with palette
• Java, 1.9, both sides
• Python, 1.9, reading only
• C, 1.9, reading only
• C, 1.11.2, writing only
• C++, 1.12.2, writing only

Old format

The following implement the previous (before 1.9) format:

• Java, 1.8
• Python, 1.4
• Node.js, 1.8