Columnar Data, from the inside out Compare the layers →

Find a concept.

Search lesson summaries, concept aliases and glossary definitions. Each lesson links to useful prerequisites and its companion evidence. Follow foundations for a first pass, then return for deeper mechanisms. Try the practice cases →

Download the lesson index · Find original evidence and reproduction recipes →

Your local reading progress

Saving is optional and off until you enable it. Only chapter links and your preferred reading path are stored in this browser. There is no account, upload, file history or record of your exercise answers.

114 chapters and 27 glossary entries match.

Glossary

Deleted row still in a file
A table can apply delete metadata while retaining a data file’s original rows. Physical file contents and visible table rows differ.
Null versus empty list
A missing list is null; a present list with no elements is empty. Both can have no leaf values while representing different records.
Why LIMIT still scans
A result limit caps returned rows. A filter, sort or aggregation may still need substantial input work to determine them.
Dictionary encoding
Repeated values can be represented by small dictionary indices. Dictionary contents and index streams are distinct.
Compression
A codec transforms encoded bytes. Its compressed size differs from encoded size, decoded arrays and process memory.
CRC checksum
An integrity check over a Parquet page body. A missing checksum makes no integrity assertion.
Bloom filter
A probabilistic membership structure whose positive answer permits false positives. A correct negative can exclude a candidate.
Page index
Optional per-page statistics and locations that can help plan selective reads. Inspecting an index alone does not execute pruning.
Field identity
A stable identifier can preserve the meaning of a field across renaming. Reusing a name does not necessarily reuse identity.
Manifest
In Iceberg, metadata that records data or delete file entries with associated statistics and partition information.
Partition evolution
Changing a table’s partition specification while historical files can retain earlier partition layouts.
Optimistic commit
Writers prepare work against a known table state and validate before publishing a new state. Conflicts can require recomputation.
Physical row position
The location of a row within a data file, distinct from a business key or its position in a filtered result.
Hash collision
Different join keys can map to the same hash bucket. The engine still needs equality checks to establish a match.
Outer join
A join that can retain unmatched input rows. Later filters can remove those rows, so predicate placement affects semantics.
IVF partition
A coarse vector-search partition used to restrict candidates. Unprobed partitions can contain exact neighbors.
Row group
A horizontal group of Parquet records, with a column chunk for each leaf column.
Column chunk
The pages for one leaf column in one Parquet row group.
Definition and repetition levels
Streams that encode missing nested values and repeated-record boundaries.
Predicate pushdown
Using a predicate earlier in a read or execution plan to avoid unnecessary work while preserving the result.
Snapshot
A committed view of a table whose referenced data and applicable deletes determine its rows.
Deletion vector
A representation of deleted row positions associated with data files. Its storage and applicability rules depend on the table format.
Checkpoint
A materialized state that provides a starting point for replay. It does not make historical data files permanent.
Selection vector
A mapping of selected row positions that can change which values an operator sees without copying every underlying value.
Arrow buffers
Typed memory buffers for values and, where required, validity, offsets and child arrays. They connect physical storage to query results.
Build and probe
Hash-join phases that construct a lookup from one input and use the other input to find candidate matches.
Recall at k
The fraction of the exact top-k neighbors present in a retrieved top-k set for a specified query and distance rule.

Chapters