HotLoop core 4.19.8 release notes
Read 4.19.4’s notes for what changed before 4.19.5, the multi-arch build of 4.19.4.
This build is amd64 only, and the next release is one small version bump that builds the arm images, the way the two runs go. stable keeps the arm images it holds in the meantime.
Also in 4.19.8
Section titled “Also in 4.19.8”-
The Logbook’s Older entries button no longer loses its page to a live refresh. The Logbook fetches the newest page again whenever the live feed says it has something new. A log line landing while an older page was on its way sent that refresh in, and the screen threw the older page away for belonging to an earlier look at the list: the button did nothing, and the click was lost. Any plant writing a log line in that moment (the window is the older request’s own time) hit it, in the old UI and the new one. The refresh now waits for the older page to land, and runs afterwards only if nothing older came. A browser test holds the older request at the network, lets a log line land and the screen refresh, then lets the page through, and wants it joined to the list; it fails on 4.19.7’s screens and passes on these.
-
Every place that picks a tag searches the whole plant, and none is cut at a number. On a tester with 2232 tags on three devices (4.19.5), the Tags screen showed 200 of them, all from the first device, the Trends picker had 201 options, a device’s Details drawer said “No tags on this device” beside a card that said 1105 tags, and searching for a tag past the first 200 said “No tags match” while
/api/tags?q=returned it. 4.19.6’s tag list changes (limit up to 5000 refused beyond, offset,X-Total-Count, the pages reading every page, windows with a count) fix those, and they are in this release. What 4.19.8 adds: the dashboard card editor’s tag fields (Gauge, Value, Trend, Map, and the Button’s writable tags) were suggestions cut at the first 3000 tags, so on a plant past 3000 a tag was never offered; they are a search the server does as the text is typed (up to 100 matches,GET /api/tags?q=&limit=100, and a newwritable=truefilter for the write pickers). A browser test with a device of 3200 tags finds a tag past 200 from the Tags search box, the device drawer, the Pin a tag picker, the Trends picker and the card editor, in the old UI and the new one.
Also in 4.19.7 (carried)
Section titled “Also in 4.19.7 (carried)”- Every HotLoop chart ships with hardware access on. A
hardware:block, three switches that default on: the app container runs privileged and as root, the node’s/devis mounted at/dev(USB radios and serial adapters, including ones plugged in later), and the node’s/run/dbusis mounted (Bluetooth through the node’s BlueZ). The pod stays on the pod network. A plain install, from the App Store or from helm, sees what is plugged into the node with no hand edit. HotLoop has no Bluetooth, Z-Wave or USB serial code of its own yet (Zigbee comes through Zigbee2MQTT over MQTT); this is the access that code, or a sidecar, will need. It costs: the pod is root on the node (the NetworkPolicy default stays on).hardware.bluetooth.hostNetworkis the opt-in, off by default, for what needs the node’s network (raw Bluetooth sockets, discovery on the node’s own network): on, the page answers on every address of the node atservice.targetPort, one HotLoop per node per port, and a NetworkPolicy does not apply to the pod. Turn any part off with the three switches; with all three off the render is, line for line, the pod 4.19.6 rendered. See Hardware in DEPLOYMENT.md. The Edge Relay chart has the same block. - The UniFi map and the Entities strip no longer write the word “null”. A script that handed a missing value to
appendorreplaceChildrenwrote it as text: a “null” above the UniFi map, and one in the Entities health strip when no node is selected. Fixed, and a browser test now fails on any text node that is only null, undefined, NaN or [object Object], on every screen.
What changed
Section titled “What changed”- The Tags screen and every device’s drawer reach every tag of every device. The page asked the server for 2000 tags and the server turned any limit over 1000 into 200, without saying so. On a plant of three devices and 2232 tags the Tags screen held 200, all from the first device, and the other devices’ drawers read “TAGS 0” and “No tags on this device”.
GET /api/tagsnow takeslimit(1 to 5000; anything else, 5001 included, is a400, never an answer with fewer) andoffset, answers the total in theX-Total-Countheader, and takesdeviceandqtogether, so a client pages to the end and knows it did. The page reads every page, draws a window of the list (500 on Tags, 200 in a drawer and in the pin picker) and says under it how many there are, with a button for the rest. The same silent ceiling was in the history lists (automation runs, alarm events, write audit, script runs, recipe applications, report runs): a limit over their ceiling was quietly the default, a tenth of what was asked. It is the ceiling now. The MCPlist_tagstool had it too, and takes the limit it is given. - The live feed sends what changed, compressed, per page. A driver reads every tag on every poll and the feed sent them all, changed or not, four times a second: about 150 KB each time on a tester’s 2232 tags. A page on an 85 KB/s link fell minutes behind, because the messages queued ahead of the current ones. The
valuesevent now carries only the readings that differ from the last one sent in value, text, unit, quality or reason, and sends an unchanged one again after 30 seconds so a steady value’s age stays honest. The stream is gzip when the browser accepts it: a tester’s first picture is about 326 KB of JSON and about 11 KB on the wire. Each page’s readings are held per tag, newest only, until its own writer can send them, and the writer rests as long as its last send took (at most five seconds), so a page on a slow link stays a send or two behind the plant, loses nothing, and takes nothing from other pages. On a link read at 300 KB/s, with every tag changing every 250 ms, a page was 27.8 seconds behind in 4.19.5 and 3.7 seconds behind in 4.19.6. - The poll counts on the device cards and the Board’s numbers tick. They were the numbers the page loaded with. They do not ride the live feed, and the refresh behind them (every 30 seconds) did not redraw anything. They are asked for every two seconds while the page is in view, one request at a time, and the screen patches its numbers in place, so a button under your hand is not replaced by it. The device drawer reads its counters from the same list.
Reading the feed
Section titled “Reading the feed”GET /api/stream sends Content-Encoding: gzip when the request says Accept-Encoding: gzip, and a script that does not say it gets plain text as before. The values event is still an array of readings in the GET /api/values shape; it just no longer repeats the ones that did not change, so a client that rebuilds its own picture from the snapshot and the events needs nothing different.
- A Kafka or Redpanda history target whose broker refuses every write for lack of disk reports “records have timed out before they were able to be produced”, and not why. The readings are not lost: the target keeps the batch, retries with a growing wait (one second up to thirty), owes the span and catches up from HotLoop’s own history when the broker answers, and its status and its system event say it is down. Redpanda refuses writes when the disk under it has less than
storage_min_free_bytesfree (5 GiB by default), and says so only in its own log (“rejecting produce request: no disk space”). The client library does not hand the refusal’s code to HotLoop in that case (it rewinds, bumps its producer epoch and retries), so there was no honest way to name the reason in this release; the design is in bead bc49eb. If a Redpanda target is down with that message, look at the broker’s log and its disk.