mxdb 0.1.13

Thin Python SDK for the mxDB bitemporal feature engine

mxDB Python SDK

mxdb is a thin Python SDK that shells out to featurectl.

For published wheels, featurectl is bundled inside the wheel and resolved automatically by MXDBClient.

Install

pip install mxdb

API Shape

The public API is namespace + entity-centric:

• client.register_feature(namespace, feature_name, value_type)
• row = client.entity(namespace, entity_name)
• row.upsert(feature_id, event_time, value, write_id=None)
• row.delete(feature_id, event_time, write_id=None)
• row.latest(feature_id, count=...)
• row.get()
• row.get_range(feature_id, date_range, disk=True)

system_time is intentionally hidden in Python. write_id is auto-generated by default, but you can pass one explicitly for idempotent retries.

Equivalent featurectl commands used under the hood:

• register <namespace> <feature_name> <value_type>
• upsert <namespace> <entity_name> <feature_id> <event_us> <value> [write_id]
• delete <namespace> <entity_name> <feature_id> <event_us> [write_id]
• get <namespace> <entity_name>
• latest <namespace> <entity_name> <feature_id> [count]
• range <namespace> <entity_name> <feature_id> <furthest> [latest] [disk|memory]

get_range Semantics

row.get_range(feature_id, date_range, disk=True) returns a newest-first list of typed values for one feature.

• date_range=(latest, furthest): returns values with event_time in [furthest, latest]
• date_range=furthest: returns everything after furthest (inclusive)
• disk=True: include values from memory + immutable segments
• disk=False: include only values currently in memory

date_range accepts epoch micros, epoch seconds (float), datetime, YYYY:MM:DD:HH:MM:SS[.ffffff], and ISO-8601 strings.

Full Example

from datetime import datetime, timezone, timedelta

from mxdb import MXDBClient

client = MXDBClient("featured.conf")

# 1) Register schema in namespace "quant"
client.register_feature("quant", "f_price", "double")
client.register_feature("quant", "f_flag", "bool")
client.register_feature("quant", "f_note", "string")
client.register_feature("quant", "f_vec", "double_vector")
client.register_feature("quant", "f_size", "int64")

# 2) Bind one entity key (row/index)
aapl = client.entity("quant", "AAPL")

# 3) Write values
base = datetime(2026, 3, 9, 12, 0, 0, tzinfo=timezone.utc)
aapl.upsert("f_price", base, 101.5)
aapl.upsert("f_price", base + timedelta(seconds=5), 102.0)
aapl.upsert("f_flag", base + timedelta(seconds=1), True)
aapl.upsert("f_note", base + timedelta(seconds=2), 'quote " ok')
aapl.upsert("f_vec", base + timedelta(seconds=3), [1.0, 2.5, 3.25])
aapl.upsert("f_size", base + timedelta(seconds=4), 1_500_000)
# optional stable write_id for retry-safe idempotency
aapl.upsert("f_price", base + timedelta(seconds=5), 102.0, write_id="price-102-v1")

# 4) Latest reads
latest_price = aapl.latest("f_price")
latest_price_history = aapl.latest("f_price", count=5)
latest_price_typed = aapl.latest_double("f_price")

# 5) Get all registered features for this entity
snapshot = aapl.get()

# 6) Range reads (latest->furthest)
bounded = aapl.get_range(
    "f_price",
    ("2026:03:09:12:00:05.000", "2026:03:09:12:00:00.000"),
)
# latest omitted => everything after furthest
open_ended = aapl.get_range("f_price", "2026:03:09:12:00:00")
# memory-only view
memory_only = aapl.get_range("f_price", (base + timedelta(seconds=10), base), disk=False)
# ISO-8601 timestamps are also supported
iso_range = aapl.get_range("f_price", ("2026-03-09T12:00:05Z", base))

# 7) Delete latest value (tombstone)
aapl.delete("f_price", base + timedelta(seconds=10))
after_delete_latest = aapl.latest("f_price")         # found=False
after_delete_history = aapl.latest("f_price", 5)     # []

print(latest_price)
print(latest_price_history)
print(latest_price_typed)
print(snapshot)
print(bounded)
print(open_ended)
print(memory_only)
print(iso_range)
print(after_delete_latest)
print(after_delete_history)

Additional Example: Multi-Entity Reads

from mxdb import MXDBClient

client = MXDBClient("featured.conf")
client.register_feature("quant", "f_price", "double")

for symbol, px in [("AAPL", 101.5), ("MSFT", 402.25), ("NVDA", 950.0)]:
    row = client.entity("quant", symbol)
    row.upsert("f_price", 1_700_000_000_000_000, px)

snapshots = {
    symbol: client.entity("quant", symbol).get()["f_price"].value
    for symbol in ("AAPL", "MSFT", "NVDA")
}
print(snapshots)

Value Types

row.upsert(...) accepts:

• bool -> bool
• int -> int64
• float -> double
• str -> string
• list[float] -> float_vector / double_vector (based on feature metadata)

Return Types

row.latest(..., count=1) returns TypedFeatureResult. row.latest(..., count>1) and row.get_range(...) return list[TypedFeatureResult] in newest-first order. row.get() returns dict[str, TypedFeatureResult].

TypedFeatureResult fields:

• found: bool
• value_type: str | None
• value: Any | None
• event_time_us: int | None
• system_time_us: int | None
• lsn: int | None

Client Methods

MXDBClient also exposes:

• checkpoint()
• compact()
• set_read_only(enabled: bool)
• backup(destination_dir: str)

Binary Resolution Order

MXDBClient resolves featurectl in this order:

1. explicit featurectl_bin= argument
2. MXDB_FEATURECTL_BIN environment variable
3. bundled wheel binary payload (mxdb/bin/featurectl or mxdb/bin/featurectl.exe)
4. featurectl on PATH

If none are found, construction fails with a clear error.