STL_SORT

STL_SORT is a scalar index that improves query performance for filters on:

Numeric fields (e.g., INT8, INT16, INT32, INT64, FLOAT, DOUBLE)
String fields (VARCHAR)
Timestamp with timezone fields (TIMESTAMPTZ)

The idea is simple: Dodil (backed by Milvus) stores the indexed field’s values in a sorted array, so the engine can use binary search instead of scanning every entity.

When to use it

Use STL_SORT when your workload frequently uses scalar filtering, such as:

Comparisons: ==, !=, >, <, >=, <=
Ranges / sets: IN
Pattern matching on strings: LIKE

Typical examples:

E‑commerce: filter by price, in_stock, rating
Observability: filter by latency_ms, status_code, error_rate
Time-based apps: filter by created_at / event_time (TIMESTAMPTZ)

Supported data types

STL_SORT supports:

Numeric fields (INT*, FLOAT, DOUBLE)
VARCHAR
TIMESTAMPTZ

How it works

Milvus implements STL_SORT in two phases:

Index build
- During ingestion, the engine collects all values for the indexed field.
- It sorts those values in ascending order.
- Each value is stored alongside its entity ID.
Query acceleration
- At query time, the engine uses binary search to locate matching values:
  - For equality, it finds the block of identical values.
  - For ranges, it finds the start and end positions, then returns everything between.
- The matching entity IDs are then used by the query executor to finalize results.

This typically reduces filtering work from a full scan O(n) to roughly O(log n + m), where m is the number of matched rows.

Create an STL_SORT index

STL_SORT does not require extra parameters.

Example: create an STL_SORT index on a TIMESTAMPTZ field named tsz.


from dodil import Client
from dodil.vbase import VBaseConfig
 
# Authenticate
c = Client(
    service_account_id="...",
    service_account_secret="...",
)
 
# Connect to your VBase database
vbase = c.vbase.connect(
    VBaseConfig(
        host="vbase-db-<id>.infra.dodil.cloud",
        port=443,
        scheme="https",
        db_name="db_<id>",
    )
)
 
# Create STL_SORT index on the field `tsz`
vbase.create_index(
    collection_name="tsz_demo",
    field_name="tsz",
    index_type="STL_SORT",
    index_name="tsz_index",  # optional
    params={},                # no extra params needed
)

What happens next

Index build may run asynchronously depending on your cluster size and data volume.
Some operations (like load_collection) may require the index to be ready.

Drop an index

Use drop_index when you want to remove the index from a collection (for example, to rebuild it with a different strategy).


vbase.drop_index(
    collection_name="tsz_demo",
    index_name="tsz_index",
)

Usage notes

Best for scalar filters: STL_SORT is meant for filtering (predicates), not for vector similarity search.
No tuning knobs: there are no algorithm-specific parameters to tweak.
Memory-mapped mode: STL_SORT does not support mmap.

If you’re unsure which scalar index to pick, start with STL_SORT for heavy equality/range filtering, then iterate based on query latency and memory usage.