Families
Families are a teaching map. They group contracts by the job they are built for. A family is coarser than a standard. The Dashboard filters by Standard, not by family. A language Overview lists that language’s roster. The Standard control lists who claims a given contract.
Theory pages cover product trade-offs. Timings live on the Dashboard.
| If you need… | Go here |
|---|---|
| Workload, standard, implementation, and how to compare | Serialization 101 |
| Full registered names and caveats | Language Overview pages |
| Timings and plots | Dashboard |
Learning goals
By the end of this page you should be able to:
- Name the families and one real example of each.
- Move from a product question to a family, and from that family to a named standard.
- Apply the comparison rule at the end of this page before treating a rank as a decision.
The families
These rows are orientation only. Real speed and size depend on the implementation and the payload.
| Family | Schema on the wire | Human-readable | Typical size | Typical speed | Cross-language | Often used for |
|---|---|---|---|---|---|---|
| JSON (text) | Optional / external | Yes | Larger | Medium | Universal | Public APIs, configs |
| Schemaless binary | Type tags / field names often present | No | Smaller than JSON | Often faster than text JSON | Wide / growing | Internal services, caches |
| Schema-driven | Numbers / layout from schema or IDL | No | Often smallest | Often fastest deserialize | Where codegen exists | Stable contracts, streams |
| Language-native | Runtime type metadata | No | Medium | Varies | Usually one runtime | Same-stack caches / graphs |
| Columnar | Table schema, often a footer or IPC schema | No | Depends on encoding and compression | Scan of few columns over many rows | Where the library can write the format | Lakes, notebooks, feature batches |
XML, YAML, CSV, and TOML sit beside the text family. When this suite gives one of them a standard id, that id is its own contract. YAML is not JSON.
Decision sketch
Work through these questions in order:
- Do people need to read or edit the payload?
- Yes → a text standard such as JSON, YAML, or TOML, where the language registers one.
- No → continue.
- Do you need a shared schema or interface description language, and evolution rules?
- Yes → a schema-driven standard.
- No → continue.
- Single language and runtime, complex graphs, and fully trusted data?
- Yes → a language-native library, only inside a hard trust boundary.
- No, but the payload really is a shared graph → the graph data set. See Data sets.
- No → a schemaless binary standard such as MessagePack or CBOR.
- Which standard inside that family? JSON, YAML, and TOML are different contracts. Protocol Buffers, Avro, and FlatBuffers are different contracts. Arrow IPC, Parquet, and ORC are different contracts. Open the Dashboard with that standard selected, then compare libraries.
Product-oriented guidance: engineering perspective.
Labels that are easy to mix
SBE (Simple Binary Encoding) sits in the schema-driven family. Its data set is columnar. Its standard is SBE. The body is word-aligned, and variable-length data sits at the end of a message or a repeating group. SBE is not a columnar file format like Parquet.
Columnar names a family and a data set, and those are different questions. Arrow IPC, Parquet, and ORC are separate standards. Arrow IPC is the in-memory interchange. Parquet and ORC are on-disk columnar files. A one-row batch pays header and alignment cost. On table_project, deserialize is the question “read one column.”
Dagr has no public specification. The Dashboard standard is No public spec. One schema, four node layouts. Packed layouts cannot store a cycle, so they are absent from the graph data set.
Graph, array, and tensor are data sets, not families. The shapes and the run configs are on the Data sets page.
Reading results fairly
Compare libraries in the same language, on the same standard, on the same data set, and on the same data type. Rank the parent row. A family is coarser than the standard id.
Schema-driven standards are often smaller and faster inside one language. That is not a universal ranking. Metrics live on the Dashboard.