Controlled software observations / synthetic arrays

Zarr archive acceptance observations

Three controlled software studies connect missing objects, fill interpretation, shard completeness and metadata age to the scientific array returned by a reader.

Metadata and rows / Full analysis / Reproduction guide / Original file hashes

Method

Generated from unchanged original observations: Zarr 3.1.3 (54 rows); Xarray 2025.10.1 and 2026.9.0 over Zarr 3.1.3 (252 rows per release); Zarr 3.4.1 strict-policy reads (138 rows). Eight routing controls and one reopen control are retained separately, not counted as CSV rows. All values are dimensionless synthetic integers or decoded missing markers.

The CSV normalizes read outcomes without replacing the originals. Each row names its source file and zero-based observation index. JSON null represents a decoded NaN in consumer values; a null values field and an error class represent a failed read. Neither missing values nor null fields are converted to zero by this generator.

Three version scopes

• Zarr-Python 3.1.3: absence, payload checksums, relocation and format-2 consolidated metadata; fill -999.
• Xarray 2025.10.1 and 2026.9.0 over Zarr 3.1.3: storage fill zero, CF masking and reduction denominators; no Dask or consolidation.
• Zarr-Python 3.4.1: format-3 strict reads, both shard-index locations and effective configuration; fill -999.

The eight API-routing controls and one reopen control are not part of the 696-row CSV. Read routing.json and observations.json for those outcomes. Different observation units must not be combined into a failure probability.

Reproduce

Use the guide and a separate environment for each dependency set. The harnesses create only private synthetic stores. Preserve the published observations and write fresh captures to new filenames. Compare observation records, not host-specific environment fields. The original source/code review is not independent peer review.

For table regeneration, retain this package at replay/public/research/zarr-archive/, then run:

node replay/public/research/zarr-archive/build-dataset.mjs --root replay

The generator verifies every selected original file against inputs.json, checks population/version boundaries and recomputes count, sum and mean. Hashes identify known bytes; trust in the expected manifest must come from outside the damaged store.

Original evidence files

• zarr-fill-evidence/experiment.py
• zarr-fill-evidence/test_experiment.py
• zarr-fill-evidence/observations.json
• zarr-fill-evidence/requirements.txt
• zarr-fill-evidence/provenance.json
• zarr-fill-evidence/sources.json
• zarr-xarray-consumer/experiment.py
• zarr-xarray-consumer/test_experiment.py
• zarr-xarray-consumer/validate.py
• zarr-xarray-consumer/observations-2025.10.1.json
• zarr-xarray-consumer/observations-2026.9.0.json
• zarr-xarray-consumer/requirements-2025.10.1.txt
• zarr-xarray-consumer/requirements-2026.9.0.txt
• zarr-xarray-consumer/provenance.json
• zarr-strict-shards/experiment.py
• zarr-strict-shards/routing.py
• zarr-strict-shards/test_experiment.py
• zarr-strict-shards/observations.json
• zarr-strict-shards/routing.json
• zarr-strict-shards/requirements.txt
• zarr-strict-shards/provenance.json
• zarr-strict-shards/claims.json

Column definitions

study

study

version

Zarr version for original/strict; Xarray version for consumer rows.

case

case

format

format

repetition

Original repeat index; null for the strict single-capture matrix.

layout

layout

state

state

selection

selection

strict

strict

sourceFile

Path relative to /research/zarr-archive/ of the original report.

sourceIndex

Zero-based index into the original report's observations array.

values

Decoded values; JSON null marks NaN in the Xarray report.

error

error

count

Count of non-null decoded values, not a completeness measure.

sum

Sum of non-null decoded values; dimensionless.

mean

Sum divided by non-null count; null for a read error.

Limits

• One Python implementation family and local stores; no field incidence estimate.
• Different experiments use different fill values and versions. Do not pool row counts into a reliability denominator.
• Strict-policy observations use format 3 in Zarr 3.4.1; stale consolidation observations use format 2 in Zarr 3.1.3.
• Xarray behavior is scoped to the two pinned releases, not current software. No Dask, concurrent writers, cloud race, crash or power-loss tests.
• Source/code self-review, not independent peer review. Input hashes identify retained bytes; they are not signatures.

Sources and reuse

CC0-1.0 for Hesela's original code and observations; third-party software and documents retain their own licenses.

1. Zarr format 2 specification: chunks and fill values
2. Zarr format 3 core specification: fill values and array reads
3. Zarr sharding codec 1.0: empty inner chunks and index layout
4. Zarr CRC32C codec 1.0
5. Zarr-Python: runtime array configuration
6. Zarr-Python 3.4.1: existing-array open and creation branches
7. Xarray 2025.10.1: open_zarr decoding contract
8. Xarray 2025.10.1: DataArray.mean and skipna
9. Xarray issue #10269: fill-mask option forwarding
10. Xarray v2026.09.0: Zarr backend implementation
11. Zarr-Python: consolidated metadata and synchronization

Definitions: Zarr, Zarr fill value, Zarr consolidated metadata.