Glossary

Quick reference for radar and infrastructure terms used across the tutorials.

Radar terms

ARCO

Analysis-Ready, Cloud-Optimized. Data already chunked, compressed, and indexed so it can be queried directly from object storage without preprocessing or downloads.

DBZH

Horizontal-polarization reflectivity factor, in decibels (dBZ). The primary radar variable — high values indicate intense precipitation, hail, or strong scatterers.

KLOT

NEXRAD radar site near Chicago, IL. The public archive at s3://nexrad-arco/KLOT is the entry-point dataset for Notebook 1.

KVNX

NEXRAD radar site near Vance AFB, OK. Used in Notebook 3 to reproduce Ryzhkov et al. (2016).

NEXRAD

Next Generation Weather Radar — the U.S. national network of WSR-88D dual-polarization Doppler radars operated by NOAA/NWS.

PHIDP

Differential propagation phase (Φ_DP), in degrees. Increases as the signal propagates through liquid precipitation; its range derivative (KDP) is used for rain-rate estimation.

Polarimetric

A dual-polarization radar variable derived from comparing horizontal and vertical pulse returns. Includes ZDR, RHOHV, PHIDP, KDP. Used to classify hydrometeor type (rain, snow, hail) and improve QPE.

QPE

Quantitative Precipitation Estimation — converting radar reflectivity into rainfall or snowfall rate using a Z–R or Z–S relationship.

QVP

Quasi-Vertical Profile — an azimuthally averaged time-height cross-section of a polarimetric variable at fixed elevation, introduced by Ryzhkov et al. (2016).

RHOHV

Co-polar correlation coefficient (ρ_HV), 0–1. High in pure liquid rain (≈ 0.98+); drops where hail, melting, or non-meteorological echoes are present.

Sweep

A single 360° conical scan at a fixed antenna elevation. A NEXRAD VCP cycles through several sweeps (typically 5–16) at increasing elevations.

VCP

Volume Coverage Pattern — a pre-defined sequence of sweeps that NEXRAD cycles through (e.g., VCP-12 general use, VCP-212 severe weather, VCP-34 clear-air mode). Each VCP has different elevation angles and scan timing.

ZDR

Differential reflectivity (Z_DR), in dB — the ratio of horizontal to vertical reflectivity. Sensitive to drop shape: large in oblate raindrops, near zero in tumbling hail or dry snow.

Infrastructure terms

DataTree

Hierarchical container in xarray (xarray.DataTree) that holds multiple datasets in a tree of nodes — the model used here to organize radar volumes by VCP and sweep.

Icechunk

ACID-compliant transactional storage layer over Zarr v3. Provides versioning, branches, and consistent reads against object storage. The nexrad-arco bucket is an Icechunk repository.

rustytree

Rust-backed xarray DataTree backend (rustytree-xarray on PyPI). Drop-in replacement for engine="zarr", ~10× faster on Icechunk repos served from S3. Used as engine="rustytree" in xr.open_datatree(...).

Zarr v3

Cloud-optimized chunked-array storage spec. Variables are split into independently compressed chunks; clients fetch only the chunks needed for a given query.

How big is the archive?

Five different numbers get called “size”, and they differ by orders of magnitude. Quoting the wrong one is the easiest way to mislead — so here is what each means for nexrad-arco/KLOT.

Logical size — what dt.nbytes reports: every array’s uncompressed extent, as if you materialized the whole archive in memory. Notebook 1 prints >100 TB for KLOT alone, and it grows with every volume the radar completes. It is the honest answer to “how much data is addressable”, and a meaningless answer to “how much storage does this need”.

Compressed source size — the bytes of the original NEXRAD Level II objects, which gzip at roughly 4:1, so their stored footprint is far below the logical size. (The ARCO chunks themselves are stored uncompressed, so for them bytes on the wire equal bytes in memory.)

Materialized size — what a native ARCO copy occupies once decoded, harmonized and rechunked. This is real duplicated storage, and it is the cost virtualization avoids.

Virtual reference size — what a repo like KLOT-lowsweeps occupies: manifests, essentially. Its chunks are byte-range references into chunks that already exist, so publishing another view of the archive costs almost nothing. Notebook 2 shows session.chunk_type(...) reporting ChunkType.virtual for exactly this reason.

Be precise about what that proves. KLOT-lowsweeps references the already-converted ARCO chunks, so it avoids a second duplication, not the first — the conversion it points into has already written a full copy of the payload. References that point straight at the raw Level II objects, skipping conversion altogether, are the stronger claim; they require a NEXRAD codec at read time and are not demonstrated on this site yet.

Bytes transferred — what a bounded query actually pulls over the network, and usually the only one a user feels. In Notebook 3’s benchmark the file-based path moves ~810 MiB compressed — about 3.2 GB once decoded — to extract the 146 MB the analysis needs; the ARCO path streams those 146 MB and nothing else. Note which ratio is which: the headline 22× is the decoded comparison; on the wire the saving is nearer 5.5×.

That last contrast is the point: the advantage is selective access, not compression. A claim like “100 TB queryable from a laptop” is true only in the logical sense — the laptop never holds 100 TB, it fetches the few hundred megabytes the query touches.