Context Engine 3.1 Release Notes
Everything that changed in Context Engine since 3.0.0: store and forward that rides out any outage and says what happened to every write, App Studio apps on MaestroHub's own address, several demo plants, and a clearer MQTT Broker page.
- Covers
- Everything after 3.0.0
- Upgrade from
- 3.0.x, with upgrade apply
Context Engine 3.1 is a minor release. It covers everything merged after 3.0.0. The notes for 3.0 itself are in the 3.0 docs: choose 3.0 in the version menu.
3.1.0 stores some 3.0 data differently, so an existing 3.0 installation runs upgrade apply once. Before you upgrade, read What Changes in 3.1 and Upgrading from 3.0 to 3.1, and see Before you upgrade below.
The themes of 3.1
- Store and forward: one write, one outcome. Every buffered write ends in exactly one place: delivered, held in its buffer, kept in Failed messages, or counted as lost with its reason. Outage reports add up, and a Kafka consumer can recognise a resent copy by the write key every record carries.
- Outages use no attempts. An outage of any length is held, and the buffer delivers when the destination is back. A message is given up, or an In order buffer blocked, only after the destination answered it with errors while it delivered other messages. Kafka's Delivery Timeout runs from the moment a produce starts, so a buffered record no longer expires before it is sent.
- Nothing is dropped silently. The first writes of a cut are buffered, a full disk is reported as no room with the right remedy, a store that can no longer be written is logged and shown as down, a write set to store and forward that its connection cannot buffer is named at every start, and a pool that falls back to its fixed size says so.
- Disk faults recover. A full data volume recovers without a restart, and a frozen data volume stalls only the store: health, sign-in and reads keep answering. Losses that could not be recorded during a freeze are written once the volume answers.
- Changing an output keeps its buffer. Editing an output's mode, delivery order, freshness or destination no longer retires its buffer; it keeps taking writes.
- App Studio apps run where MaestroHub runs (Beta). Apps are served from the address you already use, go live through one door, and read every target from their manifest, so an app moves to another line or plant by its manifest.
- Several demo plants (Beta). The Demo Plant becomes a catalog of plants, with a second, discrete-manufacturing plant beside the bottling plant.
New features and major improvements
Store and forward
- One outcome per write. A buffered write is delivered, held, kept in Failed messages or counted as lost with its reason, and outage reports show the tally: held, delivered and unaccounted, with unaccounted never above zero. See Store & Forward.
- An outage is held, a bad message is counted. Attempts count only while the destination answers other messages. See Retries: an outage is held, a bad message is counted.
- A record Kafka refuses is filed under Fix first, then replay. A record the broker marks as never retriable (for example a keyless record on a compacted topic) goes to Failed messages at once instead of being retried.
- Every Kafka record carries its write key in the header
maestrohub.write-key, so a consumer can drop a copy that an at-least-once resend wrote. See Delivery promise and Delivery promises. - Unreachable, buffering. On the Connections list and a connection's Status card, a connection in an outage whose outputs hold messages reads Unreachable, buffering instead of failed. Its tooltip says how many messages are held, from how many outputs, and how old the oldest is.
- Buffer settings. A buffer's menu item Buffer settings (formerly Failed-messages settings) sets what happens when the buffer is full: Drop oldest, Reject new or Dead-letter new. The API has its own route for it,
PATCH /api/v1/engine/durability/bindings/:pipelineId/:outputNodeId/drop-policy, guarded bydurability:write, and every change is audited. - A default pool that follows the disk. Until someone sets it, the instance-wide buffer pool is 80% of the message store's room instead of a fixed 50 GiB. The capacity page says where the number came from. See Capacity.
- Changing an output that has a buffer keeps the buffer taking writes. A change of connection removes an empty buffer at once; a buffer that still holds messages delivers them to its old connection first. See Changing an output that has a buffer.
- What In order keeps is stated exactly: the order in which runs hand values to the output. For input order from source to destination, set Run order to One at a time on every pipeline the values pass through. See Delivery order.
- Refusals under their own names.
GET /engine/durability/buffer-failurescountsdisk_floor,rejected_new,buffer_pausedandordered_wait_expiredseparately, and writes to an In order output whose run ended before their turn are now kept in Failed messages. - The instance-wide controls are a platform administrator's: the pool total, the emergency pause (which now asks for a reason) and every workspace's quota. A failed message's payload is shown only to those who may read its pipeline.
Guides
Four technical guides show store and forward at work, with the evidence behind each claim:
- Riding out a broker outage: no lost and no duplicated readings
- Sizing and protecting the buffer disk
- Keeping readings in order end to end
- Upgrading with a store-and-forward backlog
App Studio (Beta)
App Studio is a Beta feature, a paid add-on that no edition includes by default, as in 3.0. See App Studio, App Studio, start to finish and Why App Studio.
- Apps run on MaestroHub's own address. App Studio no longer needs a second host name, DNS record or certificate: apps are served under
/_apps/on the address you already use, by host name or IP, over HTTP or HTTPS, behind 443-only proxies and on sub-path installs such as Siemens Industrial Edge. Each app still runs fully isolated from MaestroHub and from other apps. Published apps need no republish, and kiosks reload themselves after updates and rollbacks. See How apps are served. - One go-live door. Approving a release, activating a version, rolling back, showcase provisioning and switching an app back on all go through one check: a signed-in person with
app:publish, the version's full reach checked against that person, and one audit row naming the approver. An AI client cannot make a version live. - A kill switch. Anyone who may edit an app can switch it off, with a reason; everyone who can switch it back on is notified, and viewers see who switched it off, when and why.
- Provable confirm. A write the manifest marks as needing confirmation is confirmed in a dialog MaestroHub draws, showing exactly what will be sent; the app cannot fake the confirmation.
- Bindings by name. App code reads every target (topics, connector functions, pipelines, dashboards) from the manifest's bindings, by name. Literal ids and paths in app code are refused at validate, preview and publish, so pointing an app at another line or workspace is a manifest change. A pipeline binding can name the Manual Trigger node the app may fire. See the Builder guide.
- A
topic_treebinding names the root of a topic subtree, so a 3D scene moves to another line or plant by its manifest. The release review lists every binding with its change against the live version. - Finding apps. The Apps page is a catalog: search by name or by what an app helps with, with the apps you can open first. An app you cannot open yet shows its listing and a Request access button, here and on its app link. Ctrl+K, search and Maestro find apps by the same rule. See Finding apps.
- Every published app has its own link. Copy app link and Open as app open the app full screen with nothing of MaestroHub around it, for a station, a bookmark or a wall; on a shared station, Sign out makes the next single sign-on ask who is signing in. See The app link.
- An app, or a whole workspace, can have its own web address. Under System Management → App addresses, a platform administrator gives an app a name such as
andon.plant.example; people who open it see only that app, with the same sign-in and permissions. The page says what DNS, the proxy, the certificate and the identity provider need. Check has MaestroHub's server open the address itself before SAML sign-in works there, and the server checks it again every day. A shared-station address can sign people out after a number of minutes without input, enforced by MaestroHub's server. Both Helm charts take the names in the newingress.extraHosts. See App addresses.
Demo Plant (Beta)
The Demo Plant is a Beta add-on for demos and education, never production, as in 3.0. See Demo Plant.
- A catalog of plants. The Overview lists the demo plants, what each makes and how it stands; with more than one plant, a switcher above every page picks the plant shown, and the sidebar has a plant switcher.
- The caliper plant. A second plant, Brightmill Precision Components, makes calipers on five operations from two CNC vendors, with its own record store and an ERP. ERP confirmations go through store and forward in order, and a quality hold uses an Approval.
- The bottling plant's use cases. An older vendor on Line 3 sends other units, and a frozen transmitter, a silent meter and a pump fault play on a loop. A stop shows its production, maintenance and finance figures, each with its source.
- A plant keeps its engine version. A plant runs the engine version it was enabled with; when a newer one exists, a banner says so, and Remove then Enable gets it.
- Remove cleans up the plant's store-and-forward buffers and Failed messages too, and a removed plant stays removed.
Unified Namespace
- The MQTT Broker page is status first, then tabs. The verdict stays at the top, with Overview, Clients, Cluster (on a cluster), Security and Settings below it. Overview shows how to connect a device and lists only the security checks that need attention. An unsaved Settings edit survives a look at another tab. See MQTT Broker.
- A stuck topic's finding says frozen, since when and after what threshold.
- Served units appear on panels, on the live channel and in the App Studio SDK.
Security and access
- Warnings for conditions that can fail to evaluate (Advanced Access Control, Enterprise). Saving a deny, a share or an access-request approval with a condition that cannot be evaluated for some people (for example a list read by position for people in no groups) answers with a warning that says who is affected, and a suggested condition where one can be offered. The Deny and Approve dialogs check the condition while you type. See Deny rules.
- A deny whose condition cannot be evaluated is held, so it refuses the person instead of being dropped.
- A credential is judged on its current state at every validation and on every transport: a revoked session, token or role takes effect on the node that revoked it before the call returns, including agent scopes, concurrent revokes and workspace deletion.
- One rule for grants and revokes. Every grant and revoke is decided by what the performer holds, and shares have one store and one revoke rule.
- Break-glass grants are marked, and a window's end takes back what it gave the person's own machine identities. See Break-glass access.
- SAML and OIDC sign-ins return only to the callback page, kept by the tab that started them, and 2.6 SAML providers keep portal sign-ins through the upgrade.
- OIDC sign-in works across several auth replicas without sticky sessions. A started sign-in is kept, sealed, in a cookie on the browser that started it, so any auth replica sharing the auth encryption key completes it, and a sign-in link opened in another browser is refused. See OIDC.
- SAML Single Logout answers the identity provider. A sign-out started at the IdP ends the user's MaestroHub sessions and is answered with a LogoutResponse (signed when Sign Requests is on). See Single Logout.
- One rule for which name a request was sent to.
X-Forwarded-Hostcounts only from a proxy listed inhttp.trustedProxies, for the links MaestroHub builds, the live-data socket, the SAML sign-in return and app addresses. - Emailed links use one configured public address.
Search and Maestro
- Search counts only what the caller may read, label facets included.
- Maestro (Beta) says on every surface why it cannot answer, worded for its reader, or stays hidden; Second Brain keeps text-only chat when the assistant did not start, and labels it.
Deployment and operations
- Enterprise (Helm) upgrades run an upgrade Job before any 3.1 service starts: it checks, waits for a confirmed snapshot when there is data to convert, scales down and runs
upgrade apply. A 3.0.0helm upgrade --reuse-valuesderives the store limit from the NATS file store. See Upgrading from 3.0 to 3.1. - Running on Kubernetes has its own architecture pages: Single-node on Kubernetes and High availability on Kubernetes. On single node, connectors reach devices under the NetworkPolicy, node-exporter clashes are refused, and the chart version equals the app version.
- The container's health check is readiness. The image runs
context-platform health, which asks/readyand prints the answer, so a container that refuses to start isunhealthywith the reason. - Re-running the native installers keeps your
config.yaml.
Notable fixes
Store and forward
- The Windows release starts again. Where the SQLite VFS that protects against a frozen volume is not built in, MaestroHub starts with that protection off and says so once at startup.
- A buffered Kafka record no longer expires before it is sent: the Delivery Timeout runs from the start of the produce, not from the record's own time.
- The first failed writes of a cut are buffered instead of being counted under
internal_error. - A destination behind a proxy that accepts and drops connections is retried at the fast pace once it answers again.
- The embedded store's cap covers the platform's own stream reservations, so a volume with little free space no longer refuses to boot.
- The outage report's held and delivered counts add up, and an always-buffered output's normal deliveries no longer count as a reconnect.
Connectors
- MariaDB queries run again, and a server that refuses a server-side timeout no longer blocks the query.
- Write forms open functions whose rows are stored as JSON.
- Executing a function needs
function:executeonly.
Platform
- A
#,?or%in the data path no longer moves every SQLite database. - The MQTT inbox reads an envelope member by member.
- The live socket opens only once a workspace is selected, and a topic subscribe is no longer logged at Info.
- A deleted workspace's OAuth2 client is announced once.
- The App Studio app list is filtered for every caller, API clients included.
- node-exporter starts on private-root hosts, and node disk figures are no longer 0.
- Demo Plant rotors spin in place in Safari, and a stop after a resume stops the plant.
Before you upgrade
- 3.1.0 needs
upgrade applyonce. It is not a swap of the image tag: if you start 3.1 on 3.0 data without the step, it refuses to start and says why. Buffered messages come through the upgrade. Follow Upgrading from 3.0 to 3.1, and read What Changes in 3.1 for who each change affects. With a backlog in the buffers, see Upgrading with a store-and-forward backlog. From 2.6.x, follow Upgrade from 2.6.x to 3.0 with the 3.1 image; it runs the 3.1 step in the sameupgrade apply. - The default buffer pool is 80% of the store's room. A pool total you set yourself is kept. An install that never set one gets the new default, which on a large disk is much more than the old 50 GiB. On the bundled Enterprise chart the store's room is its
maxSizeless the platform's own streams, so the default pool there is smaller than before: raisemaxSize(withpvc.size), or set the pool on the capacity page. See Sizing and protecting the buffer disk. - Industrial and OT connectors do not take part in store and forward, by decision. Modbus, Siemens S7, OPC UA, EtherNet/IP and similar machine protocols write directly: a write to a machine is a setpoint or a command for its moment, so a write the machine does not answer fails and the run shows it. See Industrial connectors are not buffered.
- UDP is never buffered, and a TCP write counts as sent, not confirmed. Move writes that must survive an outage from UDP to TCP or TLS. Buffers that kept the old 24-hour default retention move to keep until delivered.
- Expect
unhealthywhile MaestroHub refuses to start, and reach the upgrade page on MaestroHub's own port when a proxy routes only to healthy containers. - Stop MaestroHub before
admin-clion SQLite stores. See Stop MaestroHub before using a SQLite store. - Dashboards that read buffer-failure counts should read the new keys (
disk_floor,rejected_new,buffer_paused,ordered_wait_expired). - App Studio (Beta):
sandboxDomainandsandboxOriginare no longer read, and Helm values that still set the sandbox keys are refused with the fix; installs that kept the chart's defaults need nothing. Reload every kiosk and paired display once after upgrading. Proxies must passContent-Security-PolicyandRefererthrough unchanged on/_apps/. Apps that stored values in the browser forget them on reload. See Removed settings and Behind a reverse proxy. - A proxy that replaces
Hostand sends the browser's name only inX-Forwarded-Hostmust be listed inhttp.trustedProxies; the Helm charts, the Compose files and the Siemens Industrial Edge package keepHostand need nothing. See A proxy's forwarded name counts only from a trusted proxy. - SAML logout requests must be signed. MaestroHub refuses an unsigned, stale, repeated or misaddressed logout request from an identity provider, and logs a warning once per provider. If your IdP sends logout requests unsigned, turn on signing for them there.
- MaestroHub pages can no longer be shown inside another website's frame. Pages that embedded MaestroHub must link to it, or use the app link.
- Demo Plant (Beta): its API is per plant, under
/api/v1/demo-plant/plants/:plant/….
Get started
Download a native binary or pull the Docker image and follow Get started to have MaestroHub running in minutes.