CanViewer

Appearance

0002. canmatrix as the import adapter, not the base ​

  • Status: accepted
  • Date: 2026-11-01

Context ​

CanViewer read DBC only, with its own TypeScript parser. The owner wanted more formats, named canmatrix (Python, BSD-2-Clause) as the basis, and asked for a newer project that reads "multiple input formats, like Vector/YAML/JSON" if one existed.

A sweep of 444 searches over four angles (GitHub, package registries, community posts, YAML and JSON CAN schemas) found 52 candidates, read the top four in code, and ended with a critic pass. Nothing fit better than canmatrix. The closest alternatives each failed on a basic point: a regex DBC reader without attributes or multiplexing (omni-viewer-core), a configuration-as-code model that cannot run under Pyodide (FLYNC), no WebAssembly target (autors), a native-only library (aletheia), or no YAML or JSON database format (cantools, which stays a reference and a source of test files).

Two agents then assessed canmatrix 1.3.3: one read its code, the other converted CanViewer's own files and five opendbc files through it, on CPython and under Pyodide 314.0.7.

Decision ​

  • CanViewer's own parser stays in charge of .dbc. On every DBC point measured it was more faithful than canmatrix, it reports problems per line, and it parsed the largest test file about 3 times faster than canmatrix on CPython and 4 times faster than canmatrix under Pyodide.
  • Every other format is read by canmatrix running in the browser under Pyodide, loaded only when a non-DBC file is opened, in a Web Worker (0003).
  • canmatrix's in-memory model is mapped straight into CanViewer's DbcDatabase by a Python adapter, never through canmatrix's JSON or YAML codec or its DBC writer, both of which lose data.
  • Nothing is lost silently. What canmatrix prints while reading becomes a parse problem. CAN FD and J1939 follow CanViewer's own rules, not canmatrix's flags. Values canmatrix synthesizes, such as a signal's initial_value, are left out.
  • ARXML first, with the lazily loaded download of about 16 MB accepted by the owner; it is paid only when a non-DBC file is opened.

Consequences ​

  • Every view works for every format, because there is one model.
  • Each reader's fidelity is measured on the corpus of canmatrix's and cantools' test files and pinned, and the format reference is generated from it. Known weaknesses of canmatrix's readers (AUTOSAR 3 frames, SYM multiplexing, YAML and FIBEX reading none of the reference files) are visible there, not hidden.
  • The DBC corpus is also read through canmatrix in a test, to keep the case for the own parser measured rather than remembered.
  • canmatrix's writers are not on any data path. They matter for the backlog's export, diff and editing items, where they would need the same fidelity checks first.
  • Replacing a reader (for example cantools' ARXML reader, which reads files canmatrix fails on) is a decision for the owner.

Evidence ​

Through DBC to canmatrix to DBC, on the six test files:

  • Problems become invisible. A malformed line is caught by a bare except: that only prints. Dangling CM_ SG_ comments are dropped without a message: CanViewer's problem counts went from 17 to 0 (vw_mqb) and 7 to 0 (vw_pq).
  • CAN FD and J1939 come only from an explicit attribute on the frame; a BA_DEF_DEF_ default, a payload over 8 bytes and ProtocolType are ignored. Ford's FD count went from 331 to 330.
  • Float and double merge, the VERSION string and the network comment are not modelled, attribute values are strings, and UTF-8 files come out as mojibake unless the encoding is passed in.
  • The JSON and YAML codec drops SG_MUL_VAL_ ranges on import; its default export omits length, range, unit, values, comments and multiplexing.
  • The writers: the DBC writer emits extended-multiplexing lines with no range, the ARXML writer rounds factors to one decimal (541 of 2,752 signals in Model 3 changed), and canmatrix could not read back its own XLSX for five of six inputs.

What canmatrix does well survived: IDs and the extended flag, DLCs up to 64, signal layout, factor and offset, value descriptions, multi-line comments and real SG_MUL_VAL_ ranges.

Alternatives considered ​

  • canmatrix as the base for everything, DBC included: loses the problems per line and the measured fidelity above, and makes every DBC user wait for a 16 MB runtime.
  • Porting individual readers to TypeScript: possible (BSD-2 and MIT allow it with the notices kept), and the fallback if the download were too large. It means maintaining a reader per format; the owner accepted the download instead.
  • Converting through canmatrix's JSON: one codec for every format, but it loses data in both directions.