CSV·Parquet·Feather·HDF5의 OHLCV 데이터 벤치마크

# File-Format Storage Benchmark


# File-Format Storage Benchmark

**Environment**: the locked environment (`uv run`) — see Prerequisites below.

**Purpose**: Compare CSV, Parquet, Feather (Arrow IPC), and HDF5 on the same
1 M-row OHLCV panel along three axes — write time, read time (with forced
materialization), and on-disk size — so the trade-offs in §2.4 are
reproducible end to end.

**Learning objectives**:
1. Generate a deterministic OHLCV benchmark panel at the L scale that
   chapter §2.4 cites (100 symbols × 10,000 one-minute bars = 1,000,000 rows).
2. Time write and read for each format under one stated timing policy.
3. Force materialization on Feather / HDF5 reads so memory-mapped or lazy
   reads don't masquerade as instant.
4. Quantify the columnar-projection win (read 2 columns vs 9).
5. Render a 3-panel comparison (read time / write time / file size).

**Book reference**: §2.4 — file-based storage benchmarks.

**Prerequisites**: PyTables (the HDF5 backend) ships in the locked
environment, so `uv run python 02_financial_data_universe/20_storage_benchmark_file.py`
from the repo root is all this notebook needs — no database services, unlike
`21_storage_benchmark_database`.

> **Which environment produced the numbers**: the locked environment
> (`uv sync`, i.e. `uv.lock`), which is what §2.4 reports. The `benchmark`
> Docker image currently resolves a *newer* pandas than `uv.lock` pins, and
> pandas' newer string dtype makes PyTables store the low-cardinality
> `symbol` column about 8 bytes/row wider — enough to move the HDF5 file
> from ~71 MB to ~79 MB for the identical panel. CSV, Parquet, and Feather
> are unaffected. Run this notebook under `uv run` to reproduce §2.4.

## Setup

```python
"""File-format storage benchmark — CSV / Parquet / Feather / HDF5 at L scale."""

import gc
import os
import time
```

### Declared parameters

`BENCHMARK_SCALE` selects the panel size. The production setting is the one §2.4 quotes,
and CI overrides it to the small scale through Papermill; `ACTIVE_SCALE` and the printed
row count below say which one produced the numbers on the page.

```python
BENCHMARK_SCALE = "L"
```

`utils.storage_benchmarks` reads the scale from the environment when it is imported, so
the variable has to be set before the import rather than passed to a function afterwards.

```python
os.environ["BENCHMARK_SCALE"] = BENCHMARK_SCALE

import pandas as pd
import plotly.graph_objects as go
import polars as pl

# PyTables is the HDF5 backend; raise loudly if the image is wrong.
import tables  # noqa: F401
from plotly.subplots import make_subplots

from utils.paths import get_output_dir
from utils.storage_benchmarks import (
    ACTIVE_SCALE,
    BENCHMARK_DIR,
    N_ROWS_PER_SYMBOL,
    N_SYMBOLS,
    BenchmarkResult,
    estimate_memory_mb,
    force_materialize_pandas,
    force_materialize_polars,
    generate_ohlcv_data,
    save_benchmark_results,
    time_read,
    time_write,
    validate_result,
)
from utils.style import COLORS, show_plotly_with_alt

OUTPUT_DIR = get_output_dir(2, "storage_benchmark")
OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
```

## Timing Policy

Every number below follows one policy, applied identically to all four
formats. A comparison that mixes policies across the things it compares is
not a comparison, so this is stated up front rather than left in the call
sites:

- **Writes** (`time_write`) — a single shot, no warm-up run. Writing the
  panel is a once-per-dataset operation, so that is what we time.
- **Reads** (`time_read`) — the mean of `TIMING_RUNS` runs after one untimed
  warm-up run. **These are warm-cache numbers.** The file has just been
  written and re-read repeatedly, so it sits in the OS page cache.

Warm-cache reads are the honest description of the *repeated-access* pattern
a research loop actually has, but they flatter memory-mapped formats: Feather
maps pages that are already resident, so its read number below is close to a
best case. A first read of a cold file off disk narrows the gap to Parquet,
and on a network or object store the compressed format usually wins outright
because it moves fewer bytes. Rankings here are for warm, local, repeated
reads — the caveat that §2.4 attaches to these figures.

## 1. Generate the Benchmark Panel

`generate_ohlcv_data` returns a deterministic OHLCV panel: per-symbol random
walks under a fixed seed so format comparisons aren't muddled by data drift
across runs. At L scale that's 100 symbols × 10,000 one-minute bars =
1 M total rows, laid out over 26 regular trading sessions (390 bars each), so
the panel carries the overnight and weekend gaps real minute data has. All
symbols share the session grid, as they do in any synchronized bar panel.

```python
ohlcv_df = generate_ohlcv_data(n_symbols=N_SYMBOLS, n_rows=N_ROWS_PER_SYMBOL)
ohlcv_pandas = ohlcv_df.to_pandas()

panel_summary = pl.DataFrame(
    {
        "field": [
            "scale",
            "symbols",
            "rows per symbol",
            "total rows",
            "in-memory size (Polars, MB)",
            "in-memory size (pandas, MB)",
        ],
        "value": [
            ACTIVE_SCALE,
            f"{N_SYMBOLS:,}",
            f"{N_ROWS_PER_SYMBOL:,}",
            f"{len(ohlcv_df):,}",
            f"{estimate_memory_mb(ohlcv_df):.2f}",
            f"{estimate_memory_mb(ohlcv_pandas):.2f}",
        ],
    }
)
panel_summary
```

```python
total_rows = len(ohlcv_df)
results: list[BenchmarkResult] = []
```

## 2. CSV — Universal Baseline

CSV is the row-oriented baseline: human-readable, no compression, no schema.
Every other format is judged against it.

```python
csv_path = BENCHMARK_DIR / f"ohlcv_{ACTIVE_SCALE.lower()}.csv"

write_time, _ = time_write(lambda: ohlcv_df.write_csv(csv_path))
csv_size = csv_path.stat().st_size
results.append(BenchmarkResult("CSV", "write", write_time, csv_size, total_rows))


def read_csv_materialized() -> pl.DataFrame:
    return force_materialize_polars(pl.read_csv(csv_path))


read_time, csv_result = time_read(read_csv_materialized)
validate_result(csv_result, total_rows, "CSV read")
results.append(BenchmarkResult("CSV", "read", read_time, csv_size, total_rows))


def read_csv_columnar() -> pl.DataFrame:
    return force_materialize_polars(pl.read_csv(csv_path, columns=["close", "volume"]))


columnar_time, _ = time_read(read_csv_columnar)
results.append(BenchmarkResult("CSV", "columnar_read", columnar_time, csv_size, total_rows))
```

## 3. Parquet — Compressed Columnar Standard

Parquet's row-group layout, dictionary encoding, and Snappy compression make
it the default for analytical workloads. Column projection reads only the
row-group chunks for the requested columns.

```python
parquet_path = BENCHMARK_DIR / f"ohlcv_{ACTIVE_SCALE.lower()}.parquet"

write_time, _ = time_write(lambda: ohlcv_df.write_parquet(parquet_path))
parquet_size = parquet_path.stat().st_size
results.append(BenchmarkResult("Parquet", "write", write_time, parquet_size, total_rows))


def read_parquet_materialized() -> pl.DataFrame:
    return force_materialize_polars(pl.read_parquet(parquet_path))


read_time, parquet_result = time_read(read_parquet_materialized)
validate_result(parquet_result, total_rows, "Parquet read")
results.append(BenchmarkResult("Parquet", "read", read_time, parquet_size, total_rows))


def read_parquet_columnar() -> pl.DataFrame:
    return force_materialize_polars(pl.read_parquet(parquet_path, columns=["close", "volume"]))


columnar_time, _ = time_read(read_parquet_columnar)
results.append(BenchmarkResult("Parquet", "columnar_read", columnar_time, parquet_size, total_rows))
```

## 4. Feather (Arrow IPC) — Zero-Copy Interchange

Feather opens the file by memory-mapping it; the bare `read_ipc` call returns
almost instantly because no bytes have been read yet. We capture both the
raw open time and the time to actually materialize the columns into memory.
Only the materialized number is comparable across formats.

```python
feather_path = BENCHMARK_DIR / f"ohlcv_{ACTIVE_SCALE.lower()}.feather"

write_time, _ = time_write(lambda: ohlcv_df.write_ipc(feather_path))
feather_size = feather_path.stat().st_size
results.append(BenchmarkResult("Feather", "write", write_time, feather_size, total_rows))

gc.collect()
start = time.perf_counter()
_ = pl.read_ipc(feather_path)  # raw handle — memory-mapped, not materialized
raw_handle_time = time.perf_counter() - start


def read_feather_materialized() -> pl.DataFrame:
    return force_materialize_polars(pl.read_ipc(feather_path))


read_time, feather_result = time_read(read_feather_materialized)
validate_result(feather_result, total_rows, "Feather read")
results.append(BenchmarkResult("Feather", "read", read_time, feather_size, total_rows))


def read_feather_columnar() -> pl.DataFrame:
    return force_materialize_polars(pl.read_ipc(feather_path, columns=["close", "volume"]))


columnar_time, _ = time_read(read_feather_columnar)
results.append(BenchmarkResult("Feather", "columnar_read", columnar_time, feather_size, total_rows))
```

## 5. HDF5 — Legacy Scientific Container

HDF5 keeps a foothold in research codebases that predate Parquet. The
`fixed` format used by `pandas.HDFStore` doesn't support column projection,
so we record the columnar read as the same as the full read for fairness.

```python
hdf5_path = BENCHMARK_DIR / f"ohlcv_{ACTIVE_SCALE.lower()}.h5"


def write_hdf5() -> None:
    with pd.HDFStore(hdf5_path, mode="w") as store:
        store["ohlcv"] = ohlcv_pandas


write_time, _ = time_write(write_hdf5)
hdf5_size = hdf5_path.stat().st_size
results.append(BenchmarkResult("HDF5", "write", write_time, hdf5_size, total_rows))


def read_hdf5_materialized() -> pd.DataFrame:
    with pd.HDFStore(hdf5_path, mode="r") as store:
        df = store["ohlcv"]
    return force_materialize_pandas(df)


read_time, hdf5_result = time_read(read_hdf5_materialized)
validate_result(hdf5_result, total_rows, "HDF5 read")
results.append(BenchmarkResult("HDF5", "read", read_time, hdf5_size, total_rows))
# Fixed-format HDF5 has no column projection — record the full-read time
# so the comparison plot still has a value for the format.
results.append(BenchmarkResult("HDF5", "columnar_read", read_time, hdf5_size, total_rows))
```

## 6. Results Summary

```python
results_df = pl.DataFrame(
    [
        {
            "format": r.name,
            "operation": r.operation,
            "time_s": r.time_seconds,
            "size_mb": r.size_bytes / 1e6,
            "throughput_M_rows_s": r.rows_per_second / 1e6,
        }
        for r in results
    ]
)
```

```python
write_summary = (
    results_df.filter(pl.col("operation") == "write")
    .select(["format", "time_s", "size_mb", "throughput_M_rows_s"])
    .sort("time_s")
)
write_summary
```

```python
read_summary = (
    results_df.filter(pl.col("operation") == "read")
    .select(["format", "time_s", "throughput_M_rows_s"])
    .sort("time_s")
)
read_summary
```

```python
columnar_summary = (
    results_df.filter(pl.col("operation") == "columnar_read")
    .select(["format", "time_s", "throughput_M_rows_s"])
    .sort("time_s")
)
columnar_summary
```

```python
full_read = results_df.filter(pl.col("operation") == "read").select(["format", "time_s"])
col_read = (
    results_df.filter(pl.col("operation") == "columnar_read")
    .select(["format", "time_s"])
    .rename({"time_s": "columnar_time_s"})
)
projection_speedup = (
    full_read.join(col_read, on="format")
    .with_columns(speedup=pl.col("time_s") / pl.col("columnar_time_s"))
    .select(["format", "time_s", "columnar_time_s", "speedup"])
    .sort("speedup", descending=True)
)
projection_speedup
```

Memory-mapped reads still need to be materialized before they're useful;
the raw `read_ipc` handle time below is excluded from the comparison and
listed only so the gap to the materialized Feather read is visible.

```python
print(f"Feather raw handle (memory-mapped, not materialized): {raw_handle_time:.4f} s")
```

## 7. Visualisation

```python
fig = make_subplots(
    rows=1,
    cols=3,
    subplot_titles=["Read", "Write", "Size"],
    horizontal_spacing=0.12,
)

read_data = results_df.filter(pl.col("operation") == "read").sort("time_s")
write_data = results_df.filter(pl.col("operation") == "write").sort("time_s")
size_data = results_df.filter(pl.col("operation") == "write").sort("size_mb")

fig.add_trace(
    go.Bar(
        y=read_data["format"].to_list(),
        x=read_data["time_s"].to_list(),
        orientation="h",
        marker_color=COLORS["blue"],
        text=[f"{t:.3f}s" for t in read_data["time_s"].to_list()],
        textposition="outside",
        cliponaxis=False,
    ),
    row=1,
    col=1,
)
fig.add_trace(
    go.Bar(
        y=write_data["format"].to_list(),
        x=write_data["time_s"].to_list(),
        orientation="h",
        marker_color=COLORS["amber"],
        text=[f"{t:.3f}s" for t in write_data["time_s"].to_list()],
        textposition="outside",
        cliponaxis=False,
    ),
    row=1,
    col=2,
)
fig.add_trace(
    go.Bar(
        y=size_data["format"].to_list(),
        x=size_data["size_mb"].to_list(),
        orientation="h",
        marker_color=COLORS["slate"],
        text=[f"{s:.1f} MB" for s in size_data["size_mb"].to_list()],
        textposition="outside",
        cliponaxis=False,
    ),
    row=1,
    col=3,
)

fig.update_xaxes(title_text="Seconds (log)", row=1, col=1, type="log")
fig.update_xaxes(title_text="Seconds (log)", row=1, col=2, type="log")
fig.update_xaxes(title_text="MB", row=1, col=3)

_scale_word = {"S": "Small", "M": "Medium", "L": "Large"}.get(ACTIVE_SCALE, ACTIVE_SCALE)
print(f"Benchmark panel: {_scale_word} scale, {total_rows:,} rows")
fig.update_layout(
    title_text="Read time, write time and file size by format",
    height=400,
    showlegend=False,
    paper_bgcolor=COLORS["bg_light"],
    plot_bgcolor=COLORS["bg_light"],
    # Wider right margin so 'XX.X MB' / 'X.XXs' value labels don't crop.
    margin=dict(l=60, r=80, t=70, b=50),
)

show_plotly_with_alt(
    fig,
    "Three horizontal-bar panels, one per format in each. The left and middle plot read and "
    "write time on logarithmic axes, each bar labelled with its time in seconds and the "
    "bars sorted shortest at the bottom. The right plots file size in megabytes on a linear "
    "axis, each bar labelled and sorted the same way. The ordering of the formats differs "
    "between the three panels.",
)
```

## Key Takeaways

- **Memory-mapping is not a read.** A raw Feather handle returns almost immediately
  because nothing has been touched yet; the cost arrives when a page is first accessed.
  The materialized read is the comparison that puts all four formats on one footing, and
  it is the one plotted above.
- **No format wins all three panels.** The fastest to read, the fastest to write and the
  smallest on disk are not the same format, so the choice is a trade rather than a
  ranking, and which axis binds depends on whether the panel is written once and read
  constantly or moved across a network.
- **Column projection is what makes the columnar formats fast, and CSV cannot have it.**
  Reading a couple of columns instead of the whole schema costs roughly in proportion to
  the columns asked for on Parquet and Feather. CSV barely benefits, because reaching a
  later field on a row means parsing every field before it.
- **Compression shrinks the file, not the frame.** Parquet writes the panel at a fraction
  of the CSV size, and the in-memory footprint after read-back is identical, because both
  land in the same Arrow buffers. A format choice is a decision about disk and network,
  and it does not change what the data costs once it is loaded.
- **HDF5 fixed format is single-shot.** It can read or write the entire
  panel but offers no column projection.

### Format Picks

- **Long-term storage, cloud, cross-language**: Parquet.
- **Local interchange between Python tools**: Feather.
- **Legacy scientific Python pipelines that need append**: HDF5.
- **Human inspection or small exports**: CSV.

### Cross-References

- **Database engines** for the same panel: `21_storage_benchmark_database`.
- **Daily data lifecycle on Parquet**: `19_incremental_updates`.
- **Library-level storage primitives**: `18_data_management`.

```python
_results_csv = save_benchmark_results(results, "formats")
```
![notebook output](figures/p1_1.png)