Map(K, V)
Data type Map(K, V) stores key-value pairs.
Unlike other databases, maps are not unique in ClickHouse, i.e. a map can contain two elements with the same key.
(The reason for that is that maps are internally implemented as Array(Tuple(K, V)).)
You can use use syntax m[k] to obtain the value for key k in map m.
Also, m[k] scans the map, i.e. the runtime of the operation is linear in the size of the map.
Parameters
K— The type of the Map keys. Arbitrary type except Nullable and LowCardinality nested with Nullable types.V— The type of the Map values. Arbitrary type.
Examples
Create a table with a column of type map:
To select key2 values:
Result:
If the requested key k is not contained in the map, m[k] returns the value type's default value, e.g. 0 for integer types and '' for string types.
To check whether a key exists in a map, you can use function mapContains.
Result:
Converting Tuple to Map
Values of type Tuple() can be cast to values of type Map() using function CAST:
Example
Query:
Result:
Reading subcolumns of Map
To avoid reading the entire map, you can use subcolumns keys and values in some cases.
Example
Query:
Result:
Bucketed Map Serialization in MergeTree
By default, a Map column in MergeTree is stored as a single Array(Tuple(K, V)) stream.
Reading a single key with m['key'] requires scanning the entire column — every key-value pair for every row — even if only one key is needed.
For maps with many distinct keys this becomes a bottleneck.
Bucketed serialization (with_buckets) splits the key-value pairs into multiple independent substreams (buckets) by hashing the key.
When a query accesses m['key'], only the bucket that contains that key is read from disk, skipping all other buckets.
Enabling Bucketed Serialization
To avoid slowing down inserts, you can keep basic serialization for zero-level parts (created during INSERT) and only use with_buckets for merged parts:
How It Works
When a data part is written with with_buckets serialization:
- The average number of keys per row is computed from the block statistics.
- The number of buckets is determined by the configured strategy (see Settings).
- Each key-value pair is assigned to a bucket by hashing the key:
bucket = hash(key) % num_buckets. - Each bucket is stored as an independent substream with its own keys, values, and offsets.
- A
buckets_infometadata stream records the bucket count and statistics.
When a query reads a specific key (m['key']), the optimizer rewrites the expression to a key subcolumn (m.key_<serialized_key>).
The serialization layer computes which bucket the requested key belongs to and reads only that single bucket from disk.
When the full map is read (e.g., SELECT m), all buckets are read and reassembled into the original map. This is slower than basic serialization due to the overhead of reading and merging multiple substreams.
The bucket count can vary between parts. When parts with different bucket counts are merged, the new part's bucket count is recalculated from the merged statistics. Parts with basic and with_buckets serialization can coexist in the same table and are merged transparently.
Settings
| Setting | Default | Description |
|---|---|---|
map_serialization_version | basic | Serialization format for Map columns. basic stores as a single array stream. with_buckets splits keys into buckets for faster single-key reads. |
map_serialization_version_for_zero_level_parts | basic | Serialization format for zero-level parts (created by INSERT). Allows keeping basic for inserts to avoid write overhead, while merged parts use with_buckets. |
max_buckets_in_map | 32 | Upper bound on the number of buckets. The actual count depends on map_buckets_strategy. The maximum allowed value is 256. |
map_buckets_strategy | sqrt | Strategy for computing bucket count from average map size: constant — always use max_buckets_in_map; sqrt — use round(coefficient * sqrt(avg_size)); linear — use round(coefficient * avg_size). Result is clamped to [1, max_buckets_in_map]. |
map_buckets_coefficient | 1.0 | Multiplier for sqrt and linear strategies. Ignored when strategy is constant. |
map_buckets_min_avg_size | 32 | Minimum average keys per row to enable bucketing. If the average is below this threshold, a single bucket is used regardless of other settings. Set to 0 to disable the threshold. |
Performance Trade-offs
The following table summarizes the performance impact of with_buckets compared to basic serialization at various map sizes (10 to 10,000 keys per row). The bucket count was determined by the sqrt strategy capped at 32. The exact numbers depend on key/value types, data distribution, and hardware.
| Operation | 10 keys | 100 keys | 1,000 keys | 10,000 keys | Notes |
|---|---|---|---|---|---|
Single key lookup (m['key']) | 1.6–3.2x faster | 4.5–7.7x faster | 16–39x faster | 21–49x faster | Reads only one bucket instead of the entire column. |
| 5 key lookups | ~1x | 1.5–3.1x faster | 2.9–8.3x faster | 4.5–6.7x faster | Each key reads its own bucket; some buckets may overlap. |
PREWHERE (SELECT m WHERE m['key'] = ...) | 1.5–3.0x faster | 2.9–7.3x faster | 5.3–31x faster | 20–45x faster | PREWHERE filter reads only one bucket; full map read only for matching rows. Speedup depends on selectivity — fewer matching granules means less full-map I/O. |
Full map scan (SELECT m) | ~2x slower | ~2x slower | ~2x slower | ~2x slower | Must read and reassemble all buckets. |
| INSERT | 1.5–2.5x slower | 1.5–2.5x slower | 1.5–2.5x slower | 1.5–2.5x slower | Overhead of hashing keys and writing to multiple substreams. |
Recommendations
- Small maps (< 32 keys on average): Keep
basicserialization. The overhead of bucketing is not justified for small maps. The defaultmap_buckets_min_avg_size = 32enforces this automatically. - Medium maps (32–100 keys): Use
with_bucketswithsqrtstrategy if queries frequently access individual keys. The speedup is 4–8x for single-key lookups. - Large maps (100+ keys): Use
with_buckets. Single-key lookups are 16–49x faster. Considermap_serialization_version_for_zero_level_parts = 'basic'to keep insert speed close to the baseline. - Full map scans dominate the workload: Keep
basic. Bucketed serialization adds ~2x overhead for full scans. - Mixed workload (some key lookups, some full scans): Use
with_bucketswith zero-level parts set tobasic. ThePREWHEREoptimization reads only the relevant bucket for the filter, then reads the full map only for matching rows, giving a significant net speedup.
Alternative Approaches
If bucketed Map serialization does not fit your use case, there are two alternative approaches for improving key-level access performance:
Using the JSON Data Type
The JSON data type stores each frequent path as a separate dynamic subcolumn. Paths that exceed the max_dynamic_paths limit go into a shared data structure, which can use advanced serialization for optimized single-path reads. See the blog post for a detailed overview of the advanced serialization.
| Aspect | Map with buckets | JSON |
|---|---|---|
| Single key read | Reads one bucket (may contain other keys). All key-value pairs in the bucket are deserialized. | Frequent paths are read directly from dynamic subcolumns. Infrequent paths go to shared data; with advanced serialization, only the exact path's data is read. |
| Value types | All values share the same type V | Each path can have its own type. Paths without a type hint use Dynamic. |
| Skip index support | Works with some index types created on mapKeys/mapValues | Skip indexes can only be created on specific path subcolumns, not on all paths/values at once. |
| Full column read | ~2x slower than basic due to bucket reassembly | Overhead from Dynamic type encoding and path reconstruction. |
| Storage overhead | Minimal additional metadata | Higher due to Dynamic type encoding, path name storage, and additional metadata in advanced serialization. |
| Schema flexibility | Fixed key and value types at table creation | Fully dynamic — keys and value types can vary per row. Typed path hints can be declared for known paths. |
Use JSON when different keys need different value types, when the set of keys varies significantly across rows, or when frequently accessed keys are known in advance and can be declared as typed paths for direct subcolumn access.
Manual Sharding into Multiple Map Columns
You can manually split a single Map into multiple columns by key hash at the application level:
During insertion, route each key-value pair to the column m{hash(key) % 4}. During queries, read from the specific column: m{hash('target_key') % 4}['target_key'].
| Aspect | Map with buckets | Manual sharding |
|---|---|---|
| Ease of use | Transparent — handled by the storage engine | Requires application-level routing logic for inserts and selects |
| Vertical merge | Not supported — all buckets belong to one column | Supported — each Map column is an independent column and can be merged vertically |
| Schema changes | Bucket count adapts automatically per part | Changing the number of shards requires rewriting data or adding new columns |
| Query syntax | m['key'] works directly | Must compute the correct column: m0['key'], m1['key'], etc. |
| Bucket granularity | Per-part, adapts to data statistics | Fixed at table creation |
Manual sharding is beneficial when vertical merges are important for reducing memory usage during merges of tables with many columns, or when the number of shards must be fixed and controlled explicitly. For most use cases, automatic bucketed serialization is simpler and sufficient.
See Also
- map() function
- CAST() function
- -Map combinator for Map datatype