See how dataprof performs.

Diagnostic evidence — no established baseline. Publication mode increases the experiment budget; it does not certify precision. Warm samples share a process within each block. Separate blocks use fresh processes, but share host and OS caches. Within-run intervals do not demonstrate across-run stability; repeat-run IQR overlap is a diagnostic, not a significance test.

Optional resource collection enabled. Energy and peak RSS are per worker, including imports and warmups; warm values cover the entire block, not a single operation. Host energy zones include background load and collector overhead and are never summed. Unavailable counters remain unavailable. See the resource median/IQR tables and raw readings and paired idle baselines. Shared-runner results are collection smoke evidence, not energy-efficiency claims.

Seconds in execution order; no first sample is discarded. Process time includes interpreter startup, harness imports, adapter import/setup, operations, validation, garbage collection, IPC and exit. Import/setup is timed directly around the adapter; deferred imports remain inside operations. First operation includes initialization and, for warm workers, is the first warmup. Warmup samples are retained in the raw JSON.

First adapter import after environment preparation (preflight): dataprof: 0.009436; pandas: 0.198340; polars: 0.057147; ydata-profiling: 1.690976. Earlier installs or workflow preflights may have warmed library pages; this is not disk-cold.

Minimal-worker controls, before each round (seconds): block 1, round 1: 0.017430; block 1, round 2: 0.017411; block 1, round 3: 0.016300.

Control scope: interpreter, minimal JSON worker, IPC and exit, without harness or tool imports. It is context for combined startup costs, not a subtractable import estimate.

Measured at

2026-10-08T07:57:22.883260+00:00

Operating system

Linux-6.17.0-1022-azure-x86_64-with-glibc2.39

Processor

AMD EPYC 9V45 96-Core Processor

Samples / warmups per block

3 / 1

Process blocks

1

Experiment mode

routine

Declared host controls

GitHub-hosted runner; resource collection smoke test only

Library cache

not evicted; environment preparation and preflight may warm library pages and caches

Import preflight

imports only, in disposable workers before measurements; no fixture operations; may warm OS library pages and on-disk library caches, which are not evicted; fresh-process samples are subsequent to preflight, not first host invocations

Thread request

1

Cold file cache

warm (residency / eviction unverified)

Fixture SHA-256

1d766b0c21f3a89d78c31e76c9196471a2b5c7a87e6c39bdcff87e44b8e42272

Checkout

85c0850eb934a9f1277231a38129626297e0991c

Working tree

clean

Installed tools

dataprof 0.12.0, pandas 2.3.3, polars 1.44.2, ydata-profiling 4.18.4

Diagnostic evidence, no established baseline. Exact serialized column metrics, absence, integers, nulls and order are checked before comparison. Consumer import and profiling share one timer; lazy stream construction happens inside it. End-to-end includes both independent exports. Fresh means a new process, not cold storage.

Peak RSS includes native buffers, imports and warmups across the worker lifetime. Arrow pool observations are partial allocator evidence, not profiler-only memory.