金大哥 - deephaven-core-query-writing Skill 详情

References

Read the relevant reference before writing code — Deephaven APIs differ from similar libraries.

| Reference | Read before code that... | | --- | --- | | joins.md | uses natural_join, aj, raj, exact_join, range_join | | aggregations.md | uses agg_by, sum_by, avg_by, count_by, 20+ aggregators | | updateby.md | uses rolling/cumulative ops, EMAs, forward fill | | time-operations.md | parses/bins/manipulates timestamps; calendars, timezones | | kafka.md | consumes from / produces to Kafka | | iceberg.md | reads/writes Iceberg tables | | ui.md | builds dashboards, components, hooks, ui.table | | plotting.md | makes charts with dx | | csv.md | imports/exports CSVs, always read when importing CSV |

Core Principles

Never print() tables and never convert to pandas just to print.
Filter early. Put partition/grouping filters first in where().
In-engine over UDF. Each Python call crosses the Java boundary; prefer built-ins, java.lang.Math, and the auto-imported query-language functions.
Don't use pandas for intermediate steps — slow and unnecessary. Never use pandas unless specifically asked for.
All imports at the top. No __import__() or inline imports.

Example Table Creation

from deephaven import (
    empty_table,
    new_table,
    ring_table,
    time_table,
)
from deephaven.column import double_col, string_col

# Empty table with formulas
t = empty_table(100).update(["X = i", "Y = X * 2"])

# New table from columns
t = new_table(
    [string_col("Sym", ["AAPL", "GOOG"]), double_col("Price", [150.0, 140.0])]
)

# Ticking time table (real-time); accepts duration strings
t = time_table("PT1s")

# Ring table — bounded size, keeps last N rows
source = time_table("PT1s")
t = ring_table(source, capacity=1000)

Extracting Scalar Values

print(table) only shows the object reference. To get actual cell values from a 1-row table, drop to the Java backing: table.j_table.getColumnSource("Col").get(table.j_table.getRowSet().firstRowKey()).

Column Operations

from deephaven import empty_table

t = empty_table(100).update(["A = i", "B = i * 2", "OldName = i", "Unwanted = i"])

# only named cols, results in RAM.
# Best for: subset with expensive formula or frequent access.
t.select(["A", "B", "C = A + B"])

# only named cols, recomputed on access (no RAM).
# Best for: subset with cheap formula, sparse access, or memory pressure.
t.view(["A", "B", "C = A + B"])

# all cols + new, results in RAM.
# Best for: results that feed downstream ops (joins, aggs, further updates).
t.update(["C = A + B", "D = sqrt(C)"])

# all cols + new, recomputed on access.
# Best for: display-only columns, or when memory is a concern.
t.update_view(["C = A + B"])

# all cols + new, memoized by input value.
# Best for: few distinct inputs relative to row count (e.g. category lookups).
t.lazy_update(["C = A + B"])

t.drop_columns(["Unwanted"])
t.select_distinct(["A"])  # unique; multi-col → unique tuples
t.rename_columns(["NewName = OldName"])

Filtering

from deephaven import empty_table, new_table
from deephaven.column import string_col

t = empty_table(4).update(
    [
        "Sym = i % 2 == 0 ? `AAPL` : `GOOG`",
        "Price = 50.0 + i * 50.0",
        "Timestamp = parseInstant(`2024-06-01T00:00:00 UTC`) + i * 'P1d'",
    ]
)
filter_table = new_table([string_col("Sym", ["AAPL"])])

t.where("Price > 100")
t.where(["Sym = `AAPL`", "Price > 100"])  # list = AND
t.where("Sym.startsWith(`A`)")  # Java String methods
t.where("Sym in `AAPL`, `GOOG`")  # set membership (fast)
t.where_in(filter_table, "Sym")  # / where_not_in
t.where("Timestamp > parseInstant(`2024-01-01T00:00:00 America/New_York`)")

sym = "AAPL"  # dynamic — backtick the value
t.where(f"Sym = `{sym}`")

Joins Overview

Read references/joins.md before using joins.

Exact match: natural_join (add cols from right; NULL if no match), exact_join (errors unless exactly one match), join (all matching combinations).
Time-series: aj (closest ≤ timestamp), raj (closest ≥ timestamp).
Range: range_join — match within a range, aggregate results.
Vertical stack: from deephaven import merge; merge([t1, t2]) (column names and types must match).

Aggregations Overview

Read references/aggregations.md before using aggregations.

Single stat, all numeric cols: sum_by, avg_by, min_by, max_by, median_by, std_by, var_by, abs_sum_by — e.g. t.sum_by("Sym").
count_by("Count", "Sym") — row count per group (first arg is output col name).
first_by / last_by, head_by(N, ...) / tail_by(N, ...) — first/last row(s) per group.
weighted_avg_by / weighted_sum_by — first arg is the weight col.
agg_by([...], by=[...]) — multiple aggs in one pass (agg.avg(...), agg.sum_(...), …).
group_by / ungroup — collect into arrays / explode back. partition_by — split into sub-tables.

Dedicated aggs operate on ALL non-key columns and throw UnsupportedOperationException on String/Timestamp. Fix: .view() to keep only key + numeric first, or use agg_by with explicit cols.

Query String Syntax

Literals: Boolean true/false (lowercase). Int 42, Long 42L, Double 3.14 (no 1_000L underscores). Bare integers in / produce doubles — (year / 10) * 10 → 1987.0; fix with year - year % 10 or (int)(year / 10) * 10. String: backticks. DateTime: parseInstant(\2024-01-01T12:00:00 America/New_York`)— short aliases likeNYdo NOT work. Duration'PT1h30m'. Period 'P1y2m3d'`.

Built-ins: i row index (0-based); ii or k row key (stable).

from deephaven import empty_table

# Ternary + null handling
t = empty_table(10).update(
    [
        "Price = i * 25.0",
        "Category = Price > 100 ? `High` : `Low`",
        "Value = i % 2 == 0 ? NULL_INT : i",
        "Safe = isNull(Value) ? 0 : Value",
    ]
)

Data Import/Export

from deephaven import empty_table, read_csv, write_csv
from deephaven.parquet import read, write

t = empty_table(10).update(["X = i", "Y = X * 2"])
write_csv(t, "/tmp/o.csv")
t_csv = read_csv("/tmp/o.csv")
write(t, "/tmp/o.parquet")
t_pq = read("/tmp/o.parquet")