mergeTimelineRuns read an entry's time again on every comparison, and the
vessel layer passed a new arrow for every vessel. Times are now read once
into arrays, the leading older entries the loop would keep one by one are
sliced in one step, and the layer passes one function. The result is the
same entries in the same order; a test checks 3,000 random merges against
the previous algorithm.
Measured in the browser, one implementation per page, 9,000 vessels with
240 kept fixes each, 5 alternating pairs: 65-87 ms -> 57-61 ms per snapshot.
The label pass projects every vessel in view each time it runs. It asked
Cesium for each one, which reads the camera and allocates for every point.
createWindowProjector reads the camera's matrices and the canvas size once
and runs the same arithmetic in the same order, so the window positions are
the same numbers (a test compares them bit for bit with Cesium); outside a
3D perspective view it asks Cesium as before.
Grid cells are keyed by number, a cell's candidate is made only for its
first vessel and updated in place, and whether a name is worth a card is
kept per vessel until its name, MMSI or id changes.
Measured in the browser (perf harness, 9,000 vessels, 1 min/s): the label
pass 33 ms -> 11 ms, the forced visibility pass 42 ms -> 27 ms.
A profile of about 9,000 vessels following a playing world clock spent
~13 ms of every frame in the clock pass, moving every drawn vessel a
fraction of a pixel, and ~4 ms more re-uploading their billboards.
- Drawn vessels move again once the fastest would have moved half a pixel
on screen at the speed each last moved (measured from the world time it
was drawn at), at least once a second, and at once when the camera moves
or changes height by a tenth of its height or the view zooms. A pass that
holds them back asks for a frame, so a paused clock settles on the time.
- Hidden vessels move a share of them each pass, each within 250 ms, rather
than all of them in one frame; the visibility pass moves one coming into
view to the time first.
- A vessel's three positions are written from one surface point and normal
(Cesium's own formula) instead of three fromDegrees calls; the unused
normal is dropped. The great-circle step builds no arrays.
The pass falls from 4.6 ms to 0.37 ms a frame on average in a benchmark of
9,000 vessels (3,000 drawn) at 240 frames a second.
Playback moves something every frame, so it draws at the display's rate
rather than the parked scene's 60 FPS cap. followWorldTime reports when the
clock starts and stops playing (not live, not stepped, not waiting for data).
The top of the README now matches godseyeview.ai: official-site line, hero clip, a twelve-tile feature
grid and the video series. Docker, Codespaces, the layer table, Street Level and the voice provider
notes are folded into collapsed sections with their text unchanged. The Field Missions and Cockpit
showcase sections and the 17 launch-era GIFs are removed (68 MB of media down to under 8 MB).
package.json homepage points at godseyeview.ai.
Co-Authored-By: Claude <[email protected]>
A fix that is not a second newer (a repeat, or an older position from another
feed) never gives a course, so a source switch cannot turn an aircraft
around. The course runs from an anchor fix that moves only when a course is
read, so frequent short steps add up, and a contact reported under 2 m/s
reads none. The helper moves to src/data/courseFromFixes.js and the military
records use it as well.
Some feed rows carry neither a ground track nor a true heading for an
aircraft on the ground. Its course then fell back to 0, and an aircraft
taxiing south was drawn facing north, moving backwards or sideways. With
neither reported, the course is now read from its last fix to this one when
they are at least 30 m apart; nearer fixes are position noise and keep the
last course.
Remove the fixed 60 FPS cap. The render governor caps a parked scene at
60 ticks a second, since Cesium still runs the clock, data sources and update
listeners on every animation frame when it skips the draw, and holds
continuous render while the camera moves (moveStart to moveEnd, with pose
sampling for cameras that drift by rounding errors and never raise moveEnd).
Waits that counted frames (a tracked contact's first position, a shared
follow view, a trail's temporary copy) now count drawn time through
src/time/frameWait.js, so they last the same at any refresh rate.
Aircraft are binned into H3 resolution-4 cells (h3-js) and each cell is
drawn as its hexagon instead of a 0.5 degree square. As on gpsjam.org,
every hexagon with an aircraft is shown: the three-aircraft minimum is
gone, and discounting one bad aircraft is the only small-sample guard.
Cells are keyed by their H3 index and carry their centre and boundary.
A snapshot timeline may declare spanMs: each record of a snapshot at a
time then carries its own fixes from spanMs before the time to spanMs
after it. Flights, military flights and vessels following the clock are
placed on those fixes, so a seek or a move back within the fixes loaded
is no jump: nothing is cleared and play does not wait. The step there is
asked for once and stands in for every contact; those it has keep their
place. Fixes are merged across snapshots at least 20 s apart, at most
240 a contact, and a contact is not drawn more than a minute before its
first fix.
A weather feed that supplies its own tiles may name, per product, the
deepest level of the geographic tiling its tiles hold detail at
(frameMaximumLevel, 0-12). Global infrared is then drawn from those tiles
to that level, each given the regional infrared display transfer, rather
than from one 2048x1024 image cut to level 3. Other tiled products take
the named level in place of 6. A feed naming no usable level keeps the
layer's own detail.
Vessels out of view move on every 250 ms so one coming into view is current;
vessels at the end of their fixes ask for no frames; nothing moves while the
layer is off; 2D views judge the half-pixel step across the view's width; a
re-floor redraws vessels where they are on the next pass.
Vessels between two fixes, and live vessels a host placement places, were
redrawn at most every 250 ms, and a frame was asked for only once a second.
With nothing else asking for frames the scene drew vessels at about 1 to
2 fps live, and none at all while the world clock played at 1x.
They are now redrawn every frame the scene draws, moved in place without
new allocations. After each pass the layer works out how far the fastest
drawn vessel moved on screen and asks for the next frame for when it will
have moved another half pixel, at least once a second while any drawn
vessel moves. A vessel crossing the view close up therefore glides, and
one far away lets the scene idle. A frame already asked for later is
brought forward when a vessel speeds up.
Satellite points were propagated once a second and held still in between,
so while the world clock plays faster than real time they jumped: at
1 min/s, about 45 px once a second at globe scale.
Each SGP4 call now keeps a sample of the satellite's inertial position,
velocity, gravity and its rate of change (src/layers/satellites/motion.js).
Every frame each shown point is reckoned from its sample to the world time
and turned into the Earth-fixed frame by the sidereal angle, so points
glide at any play speed. Core satellites still take a fresh sample every
second, and at once when the world time is 30 s from it (fast play, a
seek), which keeps every point within about 10 m of SGP4. Starlink extras
take their first sample as they load and are resampled round-robin, more
per frame the faster time plays (up to 2,000 a frame); a point more than
two minutes from its sample waits for the next one. Hidden points are
moved when shown again, and the tracked satellite keeps its per-frame path.
Measured at globe scale: at 1 min/s points move on every frame, 0.4 px a
frame and at most 1.4 px (before: once a second, 45 px). With 11,880
satellites (Starlink on) the scene holds 60 fps, frames 16.7 ms median and
19.7 ms p95; reckoning costs 0.02 ms a frame for 1,342 points.
A weather row's play started the world clock even with no time to show,
where there is no timeline; it now does nothing then, as the timeline's
own Play cannot be reached. A weather layer made without a weather clock
does nothing once destroyed, so enabling it after builds no clock.
The timeline's speed follows the world clock's rate whoever set it, and
its Play runs at that rate, so it and a row's play never run at different
rates; a rate not offered shows as the nearest offered speed. Choosing a
speed sets the clock's rate first, also while a driver steps.
Marks merge on their whole label and are cut for showing after, so two
whose labels only differ past the cut stay two, and the cut counts
graphemes (code points where the runtime cannot tell them), so an emoji is
never split.
Tests that compared a DOM element with null built an assertion message
from the element when they failed, which hung the run; they compare plain
values now, so they fail at once. The timeline tests put the clock back to
real time after each.
A weather row's play ran the world clock at whatever rate it had, and
going live set that back to 1, so the speed chosen on the timeline was
lost. The world clock now keeps its rate while paused and live, and the
timeline's speed sets it whenever it is chosen, so any play runs at it.
A row's play starts from the start of the covered time when live or at
its end, as the timeline's does, and does nothing while every product the
row shows is suspended, as its reduced-motion note says.
Live, the Weather panel's frame time is the newest of each product's
frames, suspended or not, so it no longer goes away under reduced motion,
on a hidden page or without an imagery host.
A weather layer made without a weather clock builds its own when first
enabled rather than when made, so a layer never turned on leaves the
world clock alone, and its own clock (createWeatherClock({ shared: false }))
leaves the world's time where it is when destroyed; only the last shared
weather clock sends the world live.
On the timeline, a zoom or a pan clears the hovered tick's label, a label
is never wider than the track and ends in an ellipsis, and a mark's label
is cut to 60 characters (WORLD_CLOCK_MARK_LABEL_LENGTH).
The changelog entry says what a row's step does, names the earthquake
layer's new exports and says it registers with the world clock.
A layer's world clock coverage may now give each mark as an object,
{ time, label?, id?, kind? }, as well as a bare epoch ms number, and the
two may be mixed. worldClock.coverage() keeps marks as the sorted list of
distinct times, so existing readers are unchanged, and adds markDetails:
the merged marks in time order (then kind, label and id), each frozen and
holding only the fields given. Marks given alike by several sources appear
once, a bare time is dropped where a labelled mark shares it, and a mark
with no finite time is ignored. The contract is documented in the
worldClock JSDoc and docs/APPLICATION.md.
Launches label their marks with the mission's name (the launch's where it
has none), cut to 27 characters in capitals as the mission index shows
names. Earthquakes of magnitude 6 and up label theirs with the magnitude
as the map shows it and the feed's place cut to 40 characters, such as
"M6.4 · 113 km SSE of Lata, Solomon Islands"; the layer now tells the clock
when the labels change as well as the times.
On the timeline, a pointer within 6 px of a tick shows its label over the
track (nothing for an unlabelled mark), and a tap there goes exactly to
the mark's time instead of the point tapped. A tap elsewhere and a drag
work as before, and ticks stay ticks. The arrow keys still step from mark
to mark; with the handle on a mark, its labels follow the time in the
handle's aria-valuetext, and show over the track while it has focus.