NOVA vs JSON Lines: typed records, one per line
JSON Lines has only JSON’s types, and a reader cannot tell a finished stream from one that stopped early. NOVA reads JSON Lines today. NOVA Lines, a proposed stream format, keeps one record per line and the same keys, adds real types, and can end with a record count. Flat rows can also name their columns once.
events.jsonl {"ts":"2026-11-17T16:00:00Z","level":"error","service":"api","req":{"method":"POST","path":"/orders","status":502},"user":{"id":42,"roles":["admin"]},"error":{"kind":"upstream","msg":"timeout after 30s"}}
{"ts":"2026-11-17T16:00:01Z","level":"info","service":"auth","req":{"method":"GET","path":"/login","status":200},"user":{"id":7,"roles":[]}}events.novl {ts: @timestamp("2026-11-17T16:00:00Z"), level: !error, service: api, req: {method: POST, path: /orders, status: 502}, user: {id: 42, roles: [admin]}, error: {kind: upstream, msg: "timeout after 30s"}}
{ts: @timestamp("2026-11-17T16:00:01Z"), level: !info, service: auth, req: {method: GET, path: /login, status: 200}, user: {id: 7, roles: []}}
@end 2Proposed: the same shape and the same repeated keys as JSON Lines, but with real types: ts is a timestamp, level a variant. Records are appended one line at a time, and the optional @end trailer says the writer finished. A stream without it reads as open, not as complete, so tail -f still works. Open questions, spec App. C
What changes
| Topic | JSON Lines | NOVA | Spec |
|---|---|---|---|
| Appending | One value per line, append with >> | Proposed: one row or value per line in .novl; today, documents separated by --- | §4.11 |
| Field names | Repeated on every line | Repeated too; proposed: flat rows name them once in an @columns header | App. C |
| Truncation | A shorter stream is still valid | Proposed: @end <count> closes the stream; without it the stream is open | App. C |
| Types | JSON’s six kinds | Any NOVA value: timestamps, decimals, sets, variants | §2 |
| Absent fields | Leave the key out | Leave the key out, or the cell empty in a table; null is a present null | §4.9 |
| Querying | jq per line | nvf get reads JSON Lines as one list and answers in the compact profile | §7.3 |
Migrating
Today NOVA reads JSON Lines as one list. For a finished file of flat records, the compact profile writes a table with the row count filled in, so a cut-off copy fails to parse; nested records stay records. Writing .novl is part of the proposal; nvf does not do it yet. For flat rows it adds @columns, which names the top-level keys once.
nvf convert services.jsonl --to compactstdout @compat json
[2]{id,port,replicas}:
auth,8080,3
api,3000,5A cut-off table $ nvf convert services.nova --to json
× table declares 3 rows but only 2 were found (rows are indented 2 spaces)Proposed: flat rows in .novl, columns named once @columns {id, port, replicas}
auth, 8080, 3
api, 3000, 5
@end 2When to stay with JSON Lines
- You append to a log while it is being written and need a format that is specified today. NOVA Lines is only a proposal so far.
- Every tool in your pipeline already speaks JSON Lines, and you do not need types beyond JSON’s.
- You need a format that is stable today. NOVA is a 0.1 draft.
Questions
Does NOVA have a streaming format like JSON Lines?
Not yet. Today a text stream is a sequence of documents separated by --- lines. NOVA Lines (.novl) is proposed: one flow value or table row per line, an optional header with @columns, and an optional @end trailer with the record count.
How would a reader notice a stream that was cut short?
In the proposal, a writer that finishes ends with @end <count>. A reader that finds it checks the count; one that reaches the end without it reports the stream as open rather than complete. For output that is already finished, a NOVA table with [N] rows works today.
Can a record hold nested values?
Yes. For nested events such as logs, value mode is the format: each line is one flow value, as in the example above. @columns is an optimisation for flat rows and saves only the top-level keys; nested keys repeat on every line in both modes. Today the compact writer only chooses a table when every field is a scalar.
Are commas inside values a problem?
No. Rows are never split on raw commas: a cell with a comma is quoted, like any other string that needs it.