Keeping the UI thread sacred
Editora has one rule it’s held to above all others: never block the JavaFX Application Thread. An editor lives or dies on whether typing, scrolling, and highlighting feel instant, and all of those happen on that one thread. These are the rules that keep them fast.
Do the heavy work somewhere else
Tokenizing, parsing, and search all run off the UI thread on a dedicated executor, and apply their results back on the FX thread behind a generation/stale-result guard: if a newer edit has landed by the time the work finishes, the stale result is dropped. So a slow highlight pass on a big file never freezes the cursor; it just gets superseded.
Debounce and coalesce
Re-highlighting is debounced. The document overlays (whitespace markers, the
80-column ruler, the minimap, search highlights, the gutter) coalesce to one
redraw per pulse with a simple pending flag plus Platform.runLater. The
rule for new code is strict: don’t add per-keystroke or per-scroll work that
isn’t coalesced. A handful of listeners each doing “just a little” work per pulse
adds up to a slow editor.
Only touch what changed, and what’s visible
Two more multipliers:
- Incremental. An edit re-tokenizes from the changed line to the end, reusing stored per-line grammar state for the unchanged prefix, not the whole document.
- Visible-only. Overlays iterate just the visible paragraphs
(
firstVisibleParToAllParIndex … last) and skip folded lines. The rule is that nothing does O(document) work on an edit or a scroll.
There’s also a subtle one: don’t defeat JavaFX’s per-node CSS style cache. Syntax
tokens use a compound .text.<class> selector and adjacent same-style spans are
merged before being applied, so the style system isn’t asked to re-resolve more
than it must.
Know when to give up gracefully
Some files are too big to treat normally, and pretending otherwise just makes the editor janky. So there are explicit modes: at 5 MB Editora skips syntax highlighting and the minimap; at 50 MB it opens read-only with a capped load (at most the first 50 MB is read, so a multi-GB log can’t exhaust memory), and undo history is bounded. The status bar says what happened.
Stating the cost of every change
The technical rules only hold because of a habit. Every change is assessed for
its cost on the hot paths (allocation per keystroke, added FX-thread work,
extra layout/CSS passes), and that cost gets stated, even when it’s “negligible.”
When something risks a regression, I measure (temporary System.nanoTime
instrumentation) rather than guess.
Update (October 2026). These are the rules, and later releases found places that were breaking them. The posts on 0.9.7 and 0.14.0 cover the main ones. Since 0.19.0, an edit also re-tokenizes only the lines it affects instead of everything to the end of the file.