memFrame

The memFrame API is the pandas-style surface you call on a dataset context returned by an upload (or by register_tables / set_active). Every method compiles to backend-native SQL and runs in-engine across DuckDB, PostgreSQL, and ClickHouse — your data never leaves the database — and each call writes a new transient table (<table>__op_<n>) without mutating the source upload.

Every operation has a synchronous and an asynchronous twin: drop the a prefix for sync (head) and use await dataset.ahead(...) for async. Methods are dispatched lazily through the ContextManager, so you call them directly on the dataset context:

dataset = mf.upload_df(frame)

preview = dataset.head()
filled = dataset.fillna("score", method="mean")
totals = dataset.groupby("region").sum("revenue")
rolling = dataset.rolling(column="sales", window=7, func="mean", order_by="date")

dataset = await mf.aupload_df(frame)

preview = await dataset.ahead()
filled = await dataset.afillna("score", method="mean")
totals = await dataset.agroupby("region").sum("revenue")
rolling = await dataset.arolling(column="sales", window=7, func="mean", order_by="date")

Creating Datasets

mf.DataFrame mirrors the pd.DataFrame constructor: pass a dict, a list of rows (with columns), a pd.Series, or a pd.DataFrame, and it creates a new table and returns its dataset context with the full operation surface. dtype applies one dtype to every column, exactly like pandas; per-column control lives on upload_df(..., dtypes={...}).

dataset = mf.DataFrame({"name": ["Alice", "Bob"], "score": [95.5, 82.0]})
print(dataset.head(n=2))
#     name  score
# 0  Alice   95.5
# 1    Bob   82.0

dataset = mf.DataFrame([[1, "x"], [2, "y"]], columns=["id", "tag"])
dataset = mf.DataFrame(pd.Series([1, 2, 3], name="s"))
dataset = mf.DataFrame(frame, dtype="float64")

Supported dtypes: integers (int8–int64, unsigned variants) → integer columns (uint64 values past 9.2e18 overflow at insert); float32/float64, bool, datetime64 (tz-aware becomes tz-stamped); str/object/string/ category → text; decimal, date, bytes map natively. Anything else falls back to text (lossy). Empty input raises — a table needs typed columns. df.columns selects/orders columns at construction, pandas semantics verbatim; the pandas index is dropped (tables are unordered).

Domains

Domain	What it covers
Inspect	Shape, dtypes, head/tail, value counts, describe-style summaries, sampling.
Clean	Null filling (global, group-wise, forward/backward), outlier handling, type conversion, duplicate removal.
Arithmetic	Element-wise arithmetic, rounding, clipping, percentage change, column-to-column operations.
Cumulative	Running sum/product/min/max/mean/count/std/var over an ordered column.
Window	Rolling, expanding, and exponentially-weighted (EWM) operations, plus a fluent on(column) builder.
Selection	Column/row selection, loc/iloc/at, boolean and where-style filtering.
Stats	Grouped aggregations, correlation, skew/kurtosis, frequency and cross-tabulation.
Datetime	Datetime part extraction, flooring/rounding, timezone conversion, date arithmetic.
Sorting	sort_values with multi-column and mixed ascending/descending order.
Reshape	explode, melt, pivot, pivot_table, crosstab, transpose, rank.
Merge	merge/join/concat across dataset contexts on shared keys.
Transform	Feature engineering: scaling, binning, polynomial/interaction, encodings, cyclical features.

For charts, see the Plots pages. For connection, ingestion, and dataset management, see the DB Manager pages.

Return values

On a dataset context, public methods return the resulting pandas DataFrame directly. Calling the underlying wrapper classes (or a fluent builder's terminal method) returns the raw operation envelope instead — {is_error, result, new_table, ...}. Invalid operations surface as OperationError once the envelope is unwrapped.