Skip to content
GitHub

HotLoop core 4.19.4 release notes

4.19.0, 4.19.1, 4.19.2 and 4.19.3 were tagged and never published, so nothing named any of them exists and stable never moved to one. This is the first release with everything their notes list, which includes one thing that is new since 4.19.1: the charts take hostNetwork.

  • 4.19.1 added the fix for that and one more (a discovery listener that said it was up before it could hear, below) and was skipped before it released.
  • 4.19.2 added hostNetwork.
  • Both are fixed below.
  • Everything in 4.19.0’s notes (MCP, scaled writes, the store sign-in, CODESYS, the Kafka event stream, OPC UA PubSub over UADP and the fixes) is in the notes below.

What is new since 4.19.1: the charts have a hostNetwork setting.

Every HotLoop chart with a pod (the Gateway, IoT and Edge charts and the Edge Relay’s) takes hostNetwork: true. It is off by default, and with it off the chart renders exactly what 4.19.1’s did.

  • What it is for. A pod that has to hear what its node hears: UDP multicast on the node’s loopback (a PLC runtime on the same box publishing OPC UA PubSub to a multicast group over loopback), or a second network card. A pod network never carries either.
  • What it does. The Deployment gets hostNetwork: true and dnsPolicy: ClusterFirstWithHostNet, so the cluster’s DNS names keep resolving.
  • What it means for the pod. It uses the node’s addresses and ports: the page answers on every address of the node at service.targetPort, so pick a port nothing on the node uses and keep authentication on. A NetworkPolicy does not apply to a host-network pod (the cluster’s network plugin ignores them). Pin the pod to its node with nodeSelector.
  • A listener on the loopback. The OPC UA PubSub subscriber joins a group on a named card by name or by address: "interface": "127.0.0.1" in the device’s options joins on the loopback card.
  • Nothing else. The code is what 4.19.1 would have shipped. The release guard now renders every chart with and without the setting and fails if either render is wrong.

The rest of this release, from 4.19.0 and 4.19.1:

Two things in 4.19.4 change what an existing plant does, so read these before you touch a running one. Agents over MCP can write now, and the MCP endpoint always has a token. And a value you write to a scaled tag finally lands in the register the way you meant it, which means anyone who worked around the old behavior has to stop. Everything after those two is new, or a fix.

The full notes are in the README: What changed in 4.19.4.

MCP is on, it writes, and it always has a token

Section titled “MCP is on, it writes, and it always has a token”
  • Agents can write now, where an operator could. MCP is on in every HotLoop product and it reads and writes. allowMcpWrites defaults to true in the code and in all three charts. A deployment that has allowWrites on and never set allowMcpWrites gets agent writes after this upgrade, through the same gate as the screen: the user’s permissions, the tag’s writable mark and its range all still apply.
  • allowWrites did not change. It stays off by default and it still decides whether anything can write, the UI and agents alike. If it is off, nothing writes, MCP included.
  • A deployment that saved allowMcpWrites: false stays off. That includes one installed from the App Store, which stores every answer. Flip it if you want agents to write.
  • The endpoint is never open. The charts always set an MCP token. A HotLoop started without HOTLOOP_MCP_TOKEN (a Quadlet unit, a bare binary) used to serve /mcp to anybody who could reach the port and only logged a warning. It now makes a 24 byte token at its first start and keeps it in its database, so every replica and every restore carries the same one. A call without the token is a 401. A token you configured is never replaced.
  • To get the token on a bare or Quadlet install: hotloop mcp-token prints it. If you have an agent pointed at /mcp on one of those installs with no token, it stops working at this upgrade until you give it that token. That is on purpose.
  • Who made the call. A static token says which client called, never which person. With the OIDC settings on, a Vord login or a Fireball Hub assertion on /mcp makes the audit name the person (mcp:pat@example.com), and an optional write role narrows writes per caller on top of the switches above. It is off unless you set it, and the token keeps working beside it. the MCP page has the setup.
  • What was wrong. A tag with a scale, say 0.1 on a temperature register that holds tenths of a degree, shows you the engineering value: the register says 555 and the screen says 55.5. Writing went the other way without the same care. Type 55.5 for the setpoint and HotLoop put 55 in the register, which then read back as 5.5. It was the same on every screen, through the API, from an automation, a script, a recipe, a loop and over MCP, because they all go through the same write gate. The audit row said 55.5 the whole time, so nothing looked wrong in the log.
  • What changes. The number you write is in the tag’s own units, the same as the number you read. HotLoop takes the scale and the offset back off before it sends: 55.5 on a tag scaled by 0.1 puts 555 in the register, and it reads back as 55.5. A tag with no scale and no offset is not touched at all. Bits, coils, strings and Sparkplug and MQTT writes of a flag are not scaled either.
  • If you worked around it, stop. If anyone has been typing raw counts into a scaled tag to get the right number, typing 555 to get 55.5, that now sends 5550 and reads back as 555. Type the engineering value. Look at your scaled tags before you upgrade: any tag with a scale other than 1, or an offset, that somebody writes to by hand, from an automation or from a recipe is worth a look. Recipes and automations that wrote counts need their numbers changed to engineering values in the same step as the upgrade.
  • A value the register cannot hold is refused. After the scale is taken off, a count that does not fit the register, 7000 on a uint16 scaled by 0.1, is refused with a plain message that names the count and the register, and nothing is sent. It used to wrap quietly: 70000 on a 16 bit register went out as 4464. That goes for Modbus, S7, EtherNet/IP and OPC UA integer types.
  • A recipe with a value that does not fit is refused whole. Before its first write, every target is checked, so a recipe never half applies and then refuses. For Modbus, S7 and EtherNet/IP the check knows each tag’s register type from its address. For OPC UA, which is the path to a CODESYS runtime, it uses the type the tag’s last reading arrived as. A tag that has never been read since HotLoop started has no type to check against yet, so a recipe that targets it is refused at that tag’s own write, after the targets before it were written, and the application records which. Reading it once, which the poll does within a second of connecting, closes that.
  • A fraction of a count is cut off. A register holds whole counts, so 12.34 on a tenth scaled register holds 12.3, as it always did.
  • What did not change. The range check on a tag and the audit row use the number you typed. The write gate, the permissions and the refusal rules are the same.

The chart made a random admin password in a Secret the tenant can’t read, so a store install landed on a login with no way through. With the dashboard’s public key configured, a request carrying a valid signed assertion for this tenant and this Service signs that person in, makes their user the first time, keeps their role in step with what the dashboard sent, and the audit trail names them. A bad, expired, wrong tenant or wrong app assertion is refused, and local login is untouched. It is off unless a key is set, and the three product charts take the key from the store. docs/AUTH.md has the settings.

CODESYS, by the names the programmer gave the variables

Section titled “CODESYS, by the names the programmer gave the variables”

HotLoop reads and writes a CODESYS PLC through the runtime’s own OPC UA server, the way CODESYS documents: the application’s Symbol Configuration, published under the PLCopen IEC 61131-3 model. Discovery lists what the application exports with the types and the access the programmer gave them, walking into arrays and structures, and a tag is an IEC symbol that is found under whatever the project calls its PLC and application. A restart, a download or a lost link reads bad with the reason and comes back with no hands. I proved it against a real 4.22.0.0 runtime in testing running my own program, with a write the PLC’s own logic answers, and against a second program of over a hundred thousand variables, read in 2.4 seconds.

This touches a driver plants run, so plainly: the OPC UA driver asked for 200 nodes a request and never read what the server allows. A CODESYS runtime publishes OperationLimits of 100 and refuses more, and the driver answered by declaring every discovered tag a float and by reading none of the tags of a device with more than a hundred. Requests are now cut to the limits the server publishes at connect, and halved when a server that publishes none says too many. A data type that is a subtype of a standard one (a CODESYS STRING, BYTE or TIME) is followed up to the standard type, so it is typed and written as that. Every OPC UA device gets this, not only CODESYS ones.

  • What it is. HotLoop can send its readings and its system log to any broker that speaks the Kafka API: Apache Kafka, Redpanda, and the rest of that family. It is a new kind of historian target, under Settings, Historian targets, so it has the queue, the catch-up from HotLoop’s own history and the status those have. Readings go to one topic and the system log (alarms, devices going up and down, changes) to another, as JSON.
  • TLS and SASL. SCRAM-SHA-256, SCRAM-SHA-512 and PLAIN (PLAIN only over TLS), with your own CA if the broker’s certificate isn’t from a public one. Client certificates, Kerberos and OAuth logins are not supported yet.
  • It is at least once. A reading can arrive twice after an outage, because the catch-up overlaps what already went out on purpose. deviceId, tag and ts name a duplicate. Events are the live feed and have no history to catch up from, so one produced while the broker is away is lost, and the target’s status counts it.
  • A broker that vanishes mid write. The target goes down within 25 seconds and the readings in that write are owed, instead of staying up with its queue overflowing.
  • A topic made a moment ago. The first look at a topic that was just made, by you or by the target’s own last try, waits for it to exist and have a leader instead of reporting the target down.
  • What it costs. About 5 MiB on the binary of every product with a database. The Edge Relay has no historian targets and doesn’t carry it. The client is franz-go, credited in NOTICE.
  • Upgrading. Migration 0026 adds an options column to the historian targets, empty for every target that exists. Nothing changes for them. HISTORIAN-TARGETS.md has the record formats and what is and isn’t proven.
  • Subscribing. A device with the protocol opcua-pubsub listens for a publisher’s DataSets sent as UADP over UDP, unicast or multicast. A tag’s address is <DataSetWriterId>/<field index>. Variant, DataValue and RawData fields, key, delta and keep alive frames. A lost message marks that writer’s tags uncertain until its next key frame, and a silent writer reads bad.
  • Publishing. Edge and Gateway can publish the tags a layout names as UADP, off until HOTLOOP_UADP_PUBLISH_ENABLED is set. Key frames when values change and keep alives when they don’t. A bad tag is published as its status, never as its last value.
  • Not yet. Message security (signed or encrypted messages) is refused by name, as are chunked messages, discovery, events and structure fields. Delta frames are read but not sent.
  • Arrays. A field that is an array is read in Variant and DataValue messages, and in RawData when the device’s DataSet says how it is laid out ("Float[250]" for an array with its count, "Float*250" for 250 separate fields). A tag names one element, 1000/3[17], and Browse lists them.
  • Where it listens. The group, the port and the network card are the device’s own. A multicast group on a named card and a unicast port on this machine both work.
  • PROTOCOLS.md has the setup and the limits.
  • amd64 only, this once. The 4.19.4 images are built for amd64. The version tag lists amd64 alone, and stable and latest carry the new amd64 image beside the arm64 and arm/v7 images they already had, which are the previous release’s. An arm box that follows stable keeps pulling and keeps running what it ran. The next release builds every architecture again and is a version bump with nothing else in it.
  • The new screens are at /next. A rebuilt web UI is in every image behind the old one, which is still what you get at /. Try it at /next on any install; nothing about the default changes.
  • A process no longer stops the next run of a script with its own stop coming back to it. Stopping a script is announced to every process, the one that made the stop included, and the announcement heard again a moment later cancelled whichever run of that script was going then: under load a waited run was cancelled 8 ms after it began, and another never finished. A process ignores the stop it announced itself, and a stop another process announced still stops it.
  • A discovery listener is up when the broker has answered its subscriptions, and only then. A wait that ran out counted as an answer, and so did an answer on a connection that had since gone; neither does, and an answer that comes late still counts.
  • A discovery listener says it is up when it can hear. The MQTT session counted itself connected the moment the broker accepted it, and only then took its subscriptions out and waited for the broker to grant them. A listener that reported up on that sent the next message into the gap, and a device’s status or config that is not retained was lost. It reports up when the subscriptions are granted now, on the first connection and on every reconnect.
  • A pin on the board no longer wipes the board. The page only read the board when the board screen opened, so a reload on the tags screen held an empty list and the first pin replaced the saved board with that one tag. The page reads the board at start now, a pin refuses until it has, and a save carries the board’s rev, so the server answers 409 to a save over a board that changed since it was read. The rev is optional in the API: leave it out and the save is not checked.
  • A stalled OPC UA subscription is noticed. The library keeps its keep alives to itself, so a stopped publish loop looked like a quiet plant and served its last values as good. After ten publishing intervals of quiet (thirty seconds at least) the driver reads the nodes once, compares them with what it holds, and falls back to reads if any differ.
  • A PLC that goes silent with its socket open is down, not degraded. Modbus, S7 and EtherNet/IP kept their last values until TCP gave up, because the poll returned before it judged the session. They judge now. No S7 or CIP server is on the wire in testing yet, and the roadmap says so.
  • A credential typed into a URL or free text is encrypted at rest. Every screen already hid mqtt://user:pass@host and ?token= from people, but the database and the backups held them as typed. A string that carries one, by the same test redaction hides it by, is sealed whole and comes back exactly as written, so the driver still gets the URL it was given. A key turned on later seals what was stored before, and a restore asks for the key when only an address needs it.
  • Every time the API writes is UTC. A plant set to New York spelled -04:00 where one on UTC spelled Z, across about thirty fields: created_at, updated_at, an alarm’s raisedAt, an event’s time, a device’s lastPoll, a tag’s timestamp, a logbook entry, a report’s next run. They read the same on both now. If you parse these in a script that expected your local offset, it now gets Z.
  • Monthly partitions were skipped on the 31st. The history and logbook tables made their next partition with a month added to today, and on the 31st that skipped a month, so a month’s rows sat in the default partition. They start from the first of the month now.
  • The UniFi event log is ordered by when the event happened, not by when the frame got here, so a replay after a reconnect stops sitting on top.
  • A switch’s faceplate fits a phone. Eight ports were wider than a 390 pixel screen and the page scrolled sideways by 113 pixels. The ports are tight tiles now, with the PoE watts in the port’s detail, and a faceplate that still outgrows its card scrolls inside it.
  • A phone with no session and no connection gets the sign-in screen and the reason instead of an empty app, and reloads when HotLoop answers.
  • A rollback is one command. The chart index merges into the live one instead of being rebuilt from the release’s own charts, so an older version still resolves.
  • Read the first two sections above before you do anything on a running plant.
  • Migration 0026 is the only one. It is additive and every existing row keeps working.
  • A bare binary or Quadlet install gets an MCP token at its first start. hotloop mcp-token prints it.
  • helm upgrade as usual. Quadlet units follow :stable and need nothing. helm rollback to 4.18.11 is fine for this release: the database is already PostgreSQL 17 and nothing here changes it.