Skip to content

Releases

StratoWeave releases are available on GitHub.

Unreleased

Core

  • Transform actors can write their output to the layer below at any time with write_output, each write in its own transaction. Enable it with sw:direct-output under sw:transform. The transform function then gets the output written last, prev_output, and should return what the actor would write for its input, so a recompute keeps the actor's output. A write is replayed after a conflicting commit. fleetmgr.cfs.UpgradeCampaignTransform is an example. #698 #783
  • A transform's output replaces the configuration it wrote before, instead of being applied as a diff. Fixed source deletion, concurrent edits, restored list membership and subscriptions on newly committed list elements. #680
  • A full configuration replacement, like a transform's output, skips list elements and container children that are the same node as in the last replacement committed from that source. They get no transaction, lock or commit, and their transforms don't run. An equal value in a newly built node is not skipped. #731
  • A session commit reports done and complete only after every transform below it has committed. Before, a link subscriber could commit without its linked input. #685
  • Fixed a deadlock in update_dynstate when the root transform was blocked on a get_data call to the same transaction. #675
  • Optional parameters declare explicit defaults, which newer Acton versions require. The transform generator emits them too.
  • Child processes, environment variables and stdin go through the Env services env.process, env.vars and env.stdio. #704
  • Fixed the etcd store leaving a request without a result when etcd answers with an HTTP error whose body is not UTF-8. #724
  • Fixed subscription data running in the actor of the device adapter or TTT layer that delivered it, when the consumer passed its own actor method. declare_subscriptions() of DeviceMgr, ttt.Layer and adapters.adapter.DeviceAdapter takes the data callback, and the layer's delta, as an action instead of a proc. #763
  • A software adapter has a static prepare_estimate(): the seconds the copy of the image takes, for planners. The IOS XE adapter gives 130. The unused estimate(staged) is removed. #784
  • Reads of a transform's data no longer copy each list entry to add its keys when the entry already has them. In Fleet Manager with 10,000 devices, a read of the fleet configuration takes 37 ms instead of 45 ms. #815
  • Acton library modules such as json, http and argparse are imported from std, as Acton moves them out of base acton#3410. #833

Northbound APIs

  • Fixed NETCONF <get-config> returning the operational state published by transforms. #683
  • RESTCONF supports the content query parameter on GET. #686
  • RESTCONF supports the fields query parameter on GET, which narrows the reply to the selected nodes, like GET /restconf/data/demo-fleet:device=cpe-16?fields=software(target;state). The reply for a list entry keeps its keys. /restconf-state/capabilities lists the fields capability. An invalid fields value, fields on any method other than GET, and a query parameter given twice get 400 Bad Request. #805
  • Fixed RESTCONF GET, NETCONF <get> and yang-push filters that could not reach an operational subtree under a config node, like /software:software/upgrade-campaign=foo/state in fleetmgr. #677
  • Fixed /restconf-state/capabilities listing YANG modules next to the RESTCONF protocol capabilities. #676
  • RESTCONF replies use compact JSON and XML. ncurl output and the /device/<name>/target, /running and /diff inspection endpoints are still indented. #684
  • The /layer/<idx> inspection endpoint takes the same content query parameter, so it can return a layer's operational state too. #736
  • The NETCONF server serves the YANG library at /yang-library and /modules-state. Besides the modules of the CFS schema, it lists ietf-subscribed-notifications and ietf-yang-push as implemented, with the features subtree, xpath and on-change. The <hello> has one yang-library:1.1 capability instead of a capability per module. RESTCONF serves /modules-state too. #716
  • The NETCONF server serves <get-schema> (RFC 6022) for each module and submodule of the CFS schema, of ietf-yang-library, and of the modules that the server implements itself and their imports. The YANG library lists ietf-netconf-monitoring as implemented. #820
  • The NETCONF server serves RFC 5277 create-subscription and advertises the :notification and :interleave capabilities. Application code sends a notification to the subscribed sessions with publish() on NetconfServer.event_hub(). There is no replay yet, and a session with an RFC 5277 subscription cannot also hold RFC 8639 subscriptions. #752
  • The /device/<name>/modules and /device/<name>/info inspection endpoints give the adapter of each module. #756
  • A NETCONF <edit-config> transaction is complete once it is committed to TTT and persisted. Device interaction is left outside the transaction, so the edit no longer waits for the devices to apply their configuration. Before, a device that never connected kept the transaction from ever completing, and since a Fleet Manager top sends one edit at a time to each flotilla, a single such device blocked all further configuration to its flotilla. #792
  • The HTTP server sends embedded static assets unchanged as bytes, so an application's web UI can include images, fonts and other binary files, which are served with their content type. stratoweave.main() takes static_assets as dict[str, bytes] instead of dict[str, str].
  • The HTTP server compresses text responses, such as RESTCONF and HTTP API replies and the web UI's files, with gzip or deflate when the client accepts it in Accept-Encoding. In Fleet Manager with 100,000 devices, the 14 MB fleet list goes out as about 1.3 MB. #832

Devices & Adapters

  • SNMP device adapters are generated from MIBs. The system generator translates the MIBs to YANG and generates Acton types and subscription paths from them. Transforms choose the polling period, and the SNMPv2c adapter uses credentials/password as the community. Mini has an example ietf_snmp device type with SNMPv2-MIB and IF-MIB. See the SNMP devices reference. #701
  • Device and software adapters get the Env instead of a TCPConnectCap, so they can use other transports, like UDP for SNMP, or start subprocesses. #702
  • Each device publishes its operating system family, release, chassis model and serial number under state/platform. A platform probe reads them on every connect and after a software install. There are probes for IOS XE, IOS XR, Junos, SR Linux, SR OS and ArcOS, and an app sets its probes with platform_probes in its device_types.act. See the device platform reference. #712
  • Transforms can subscribe to a device's YANG event notifications with declare_notif_subscriptions() on the DeviceMgr, naming them in NotifSpecs. Each NotifEvent carries the event time, the payload decoded against the device schema and the raw XML. The NETCONF adapter uses RFC 8639 establish-subscription on the main session where the device supports it, and otherwise an RFC 5277 subscription on a session of its own. Delivery is at most once, with no replay. Pass a synced action to learn when the subscriptions are in place, before sending an RPC whose result arrives as an event. See the event notifications reference. #752
  • ncurl notifications prints the event notifications a device sends, all of them or those named with --notification module:name, and how the adapter subscribed to them. #752
  • When a device advertises ietf-netconf-notifications (RFC 6470), the DeviceMgr subscribes to netconf-config-change and resyncs the device's running configuration when one arrives. A change made outside StratoWeave is found without waiting for a reconnect. StratoWeave's own commits cause a resync too, because the DeviceMgr does not check changed-by. #753
  • adapters.iosxe_cli.IosXeCliAdapter runs IOS XE commands over SSH, for what NETCONF lacks on a device, or on older releases where NETCONF does not start at all. It returns what it reads in the stratoweave-iosxe-cli YANG module: the files on a file system from dir, the release, model, serial number and boot image from show version, and the output of show install summary and show bootvar. A device type that uses it adds adapters.iosxe_cli_yang.models to its schema. #755
  • A device can run extra adapters beside the adapter of its device type, for what that adapter does not have, like a CLI adapter for an RPC that the device lacks over NETCONF. An extra adapter is a device type too, listed in the device's extra-adapter list with an optional address, port and credentials. Without a port it uses its own default port, not the port of the address. A module goes to the first connected adapter that has it, the adapter of the device type first, and moves when an adapter connects or loses its connection. Gets, subscriptions, RPCs and event notification subscriptions go to the adapters of the modules that their top-level nodes name, and configuration goes only to the adapter of the device type. state/adapter shows whether each adapter is connected. An adapter runs as an extra adapter when it passes its device type to the DeviceMgr's on_connect() and on_disconnect(), as the NETCONF, SNMP and IOS XE CLI adapters do. See the device management reference. #756 #765
  • DeviceMgr.get_connected_modules() gives the modules of a device's connected adapters, and get_unconnected_durations() how long each adapter that is not connected has been without a connection. #799
  • Fixed devices without yang-library or netconf-monitoring, like Junos 10.4, getting an empty module set, so transforms computed nothing for them. Their module set has the other capabilities of their <hello>, named by URI, except the NETCONF protocol ones. The Junos platform probe applies on the http://xml.juniper.net/netconf/junos/1.0 capability. #721
  • The NETCONF adapter fails a get whose filter names a node that the device schema lacks, with Filter node <path> not found in device schema. Before, those nodes were dropped from the reply without an error. #712
  • ncurl takes the repeatable --ssh-key-exchange, --ssh-host-key-algorithm, --ssh-cipher and --ssh-mac options for devices that only offer old SSH algorithms, like Junos 10.4. Its new platform command prints the platform the probes read. The default port is 830 instead of 42830, and without --password it prompts for the password instead of using admin. #717
  • After refused credentials, an untrusted host key or no common SSH algorithm, the NETCONF adapter tries to connect again once a minute instead of every second. Repeated failed logins can lock the account on some devices. A change to the device's settings connects at once. #720
  • Fixed NETCONF RPCs to Junos 10.4 waiting until their timeout. Its replies carry xmlns twice, which is not valid XML. The NETCONF client drops the repeated attribute with the new duplicate-attributes fixup. #719
  • Fixed "Invalid NETCONF chunk framing" errors around NETCONF session restarts, seen with large mock fleets. After such an error the device adapter restarts the session once, not once for every write. #691
  • Fixed the mock NETCONF server keeping a session's yang-push subscriptions after the session ended. Every client restart added another set of updates. #692
  • Fixed a transform's subscriptions to a device being lost when they were declared before the DeviceMgr had its adapter, for example when the device and the transform's input are written in one transaction. The DeviceMgr passes them on to the adapter it gets later. In Fleet Manager, a campaign could keep upgraded devices pending because the top never subscribed to their flotilla. #747
  • The NETCONF adapter logs in with the first private key in the device's credentials/key, with or without a password, on the main session and on the RFC 5277 notification sessions. Before, it connected only with a password. ncurl takes a private key file with --key. #776
  • An extra adapter can have SSH settings of its own in extra-adapter/ssh, which replace the device's ssh for that adapter, as credentials does. The IOS SSH server of a C1111 sends no server-sig-algs, so an RSA key needs public-key-algorithm ssh-rsa on its CLI port, while its NETCONF port takes the RSA SHA-2 signatures. #776
  • The verify RPC of the IOS XE CLI adapter runs the IOS XE verify command in its forms. Without an algorithm, the device checks the hashes and the signature that an image holds. With md5 or sha512, it computes the digest, and with a hash too, it compares them and verified says whether they match. #755
  • The IOS XE CLI adapter runs install commands with its install RPC: add, activate with an abort timer, commit, abort, remove and rollback. It says whether the device reloads. A device in bundle mode refuses install add alone, and install add file F activate commit converts it to install mode. #777
  • The configure RPC of the IOS XE CLI adapter types configuration lines in configure terminal over an SSH shell, one at a time, and stops at the first line that the device rejects. write-memory saves the configuration. Install mode boots bootflash:packages.conf, which the boot statement has to name, and a conversion sets it this way. #777
  • In mock mode, the IOS XE CLI adapter answers from a CliMock, which a test gets with get_mock() and gives the replies of the device. #777
  • The IOS XE CLI adapter logs in once per device and keeps an SSH shell open, in which it types one command at a time. IOS XE closes the connection when an exec channel ends. The device is connected while the shell is ready. When a login fails, the connection closes, or a command gets no prompt within its time, the adapter reports the device as disconnected and logs in again. After a minute without a command, it types an empty line, so a dead connection shows. #765
  • The install RPC of the IOS XE CLI adapter answers the question that install remove asks before it deletes files, which has no prompt-level none on IOS XE 17.15 and comes twice on 16.8, and the question of install abort. At debug level the adapter logs each command that it types and each question that it answers. #780
  • The ios-xe-cli platform probe reads show version through the IOS XE CLI adapter, after the NETCONF probe, so an IOS XE device without NETCONF, like 16.8 on a C1111, has a platform. Probes run over the modules of the connected adapters, so a device whose NETCONF went down reads its platform over the CLI. #799
  • The IOS XE CLI adapter has a copy RPC, which fetches a file onto the device and tells how far it is from the ! marks of the copy, and a delete RPC. It reads when the running and the startup configuration last changed, and the index of each file in dir. #799
  • The IOS XE CLI adapter reads the BGP neighbors into Cisco-IOS-XE-bgp-oper from show bgp ... neighbors, in the shape that NETCONF gives them, and realizes subscriptions by polling, as the SNMP adapter does. The fleetmgr BGP upgrade checks then work on a device without NETCONF. The polling moved out of the SNMP adapter into the adapters.polling.Poller actor, which both adapters use. #799
  • Fixed event notifications of the IOS XE -events-oper modules, like interface-admin-state-change in Cisco-IOS-XE-ios-events-oper, never arriving. The NETCONF adapter subscribed to them on the RFC 5277 NETCONF stream, which the device accepts but sends nothing on. It now subscribes to them on yang-notif-native, as it does for the -events modules. #782
  • The device setting telemetry selects how the adapter receives the data of subscriptions: push (YANG-push) or poll. With poll, the NETCONF adapter reads with periodic <get> and receives event notifications through RFC 5277. Without the setting, ArcOS devices are polled: ArcOS refuses every other RPC on a session that holds a subscription.
  • Runtime schema fetch no longer skips tailf-rollback. It was on the list of broken modules because acton-yang failed to compile it, and tailf-netconf-with-rollback-id imports it, so a device that lists both, like ArcOS 8.2.1A, got no schema at all. #837
  • The device platform has last-boot: when the device last booted, in seconds since the Unix epoch by the device's clock. The IOS XE probe reads it from boot-time and the SR Linux probe from last-booted; the other probes leave it out. The IOS XE mock boots when it starts and at each reload of an upgrade. [#841]
  • Fixed a NETCONF resync or checked edit that never finished when the device configuration does not match the schema, like a value outside a range. The resync now fails with the decode error. A checked edit fails with it before it edits the device. When only the configuration after the commit does not decode, the edit succeeds and the next checked edit resyncs first.
  • A DeviceMgr reads the running configuration of its device, and subscribes to netconf-config-change, only from the first configuration for the device, the first get_config() or resync(), or from the first connect with track_config=True. Before, it read it on every connect, so tools that only read data or subscribe, like ncurl monitor, read the whole configuration too.
  • ncurl cli is an interactive command line in the style of the Junos CLI. Its commands, completion with TAB and space, and ? help come from the device schema, shaped by the Junos cli-feature and the Tail-f cli-drop-node-name and cli-no-keyword statements in it. show configuration prints the running configuration in curly braces, or in other formats with | display set, | display xml and | display json. configure starts configuration mode on a candidate kept in the CLI, with set, delete, edit, up, top, rollback, run, show and show | compare. With a command as arguments, it runs that command and exits. Without a terminal, it reads commands line by line from stdin.

Software Images & Upgrades

  • Software images can be uploaded over HTTP into an image folder, set with --software.image-dir. Upload with PUT /image-file/<name>, list with GET /image-file and remove with DELETE /image-file/<name>. The new stratoweave-images module adds an image list, /images/image. Add swyang.images to the top layer to use it. The list and the folder are not kept in sync yet. #703
  • An image in /images/image has a family, like c8000v-universalk9, and a default flag. A family has at most one default image. /images/url-base says where devices fetch the images from. #742
  • The device software state has two new leaves: stage, the step the run is in or ended in, and reason, set when a run fails or rolls back. The IOS XE adapter takes reason from the txn-sub-state of the failed install. #669
  • The IOS XE mock can serve a BGP neighbor for the upgrade checks. Two new mock failures use it: with bgp-down the pre-check refuses the install, and with bgp-lost the post-check fails and the device rolls back. #667 #678
  • The IOS XE mock serves seven interfaces in Cisco-IOS-XE-interfaces-oper and two full OSPF neighbors in Cisco-IOS-XE-ospf-oper for the upgrade checks. Two new mock failures change them while the activated release runs, until a commit or an abort: interface-lost takes the line protocol of GigabitEthernet0/0/1 and Virtual-Access2 down, and ospf-lost keeps the OSPF neighbor 10.255.0.3 in init. #821
  • Removed running-release from the device software state. The release a device runs is in state/platform/version, which is read again when a software install ends, also after a rollback. #718
  • Mock devices set the platform that the IOS XE mock reports under mock/platform: version, the release before an upgrade, model and serial-number. Without them a mock runs 17.18.02 on a C8000V. mock/platform/version replaces mock/software/running-release. The mock reports a release the way a device does, 17.18.02 as 17.18.2. #718
  • Mock upgrades vary in length. Each takes a time within mock/software/upgrade-duration-variance percent of upgrade-duration, either way, 20 by default, so devices released together no longer all finish together. 0 keeps every upgrade at exactly upgrade-duration. #770
  • A software veto holds only the install. The software manager still reads the release the device runs and reports up-to-date, or the new upgrade-needed when it differs. The vetoes are read again right before the install, so a veto added during the pre-checks holds it too. Before, a veto stopped even the read. #775
  • The software state says which campaign and target release its status answers, in state/upgrade-campaign and state/target-release. A final status stands, so a reconnect or an unrelated config change no longer runs a failed upgrade again. The install installs the release the run read the device for, and a target changed during the pre-checks is read again before anything is installed. #775
  • Under a veto, a read of the release that fails gives unknown with the error, and is read again when the device connects, or after a while. With runtime-schema-fetch, a NETCONF get or RPC on a device that has not connected yet fails as not connected, instead of with a missing schema. #775
  • With allow-staging, the software manager prepares the target while a veto holds the install: it copies the image to the device ahead of the window, once per intent. The result is in state/job[prepare], which also says which release it is for. A failed prepare leaves the status upgrade-needed. The prepare of a run reports in the same entry. Before, allow-staging was stored and not used. #775
  • An image in /images/image can have the md5, sha256 and sha512 digests that its vendor publishes in its checksum list, and its size in bytes. A campaign copies them with the image URL into the software intent of each device, as software/image-checksum and software/image-size, for the device's adapter to check the staged image against. #778
  • The IOS XE software adapter picks how each step of an upgrade runs from the modules of the device's connected adapters: over NETCONF where the device has the module for the step, else over the IOS XE CLI adapter. That covers reading the release and the versions, the transfer, the check, add, activate, commit and abort. It lists bootflash: over NETCONF (Cisco-IOS-XE-platform-software-oper) or the CLI, and checks a staged image with the CLI adapter's verify: against the image checksum of the intent, sha512 before md5, or else against the signature that the image holds. A device that can compute no checksum compares the size of the staged file with software/image-size. #781
  • IOS XE prepare and install can run again at any time and do only what is left. A release that the device has added needs no image file, and a staged file that passes its check stays. A missing or failing file is transferred and checked again. Install activates a release that the device has added, and a release that the device runs activated and not committed is committed instead of reported up to date. When install add refuses an image as bad, the adapter deletes the file over NETCONF, so the next run transfers it again. #781
  • Before it stages an image, the IOS XE adapter waits until each adapter of the device that is not connected, such as a CLI adapter that connects after NETCONF, has been without a connection for 2 minutes. The wait counts from when the adapter lost its connection or was set up, so the later stages of an upgrade do not wait again. #781
  • The IOS XE software adapter converts a device in bundle mode, which it also detects over NETCONF, to install mode, and on IOS XE 16 it installs with one command, install add file ... activate commit. A C1111 went from 16.8.1 in bundle mode to 17.15.06 over the CLI alone. #799
  • Before a transfer the IOS XE adapter looks for room for 2.2 times the image size, and before an add for 1.2 times. When short, it removes the software that the device does not run, once. The manager calls cleanup after the commit. A transfer that failed for lack of room says so. #799
  • The IOS XE adapter removes another added release before it stages the target, and aborts another release pending commit before an install. It refuses to install while a change that someone made to the running configuration is not saved, and saves a change that the device made itself as it booted. #799
  • The IOS XE commit window comes from the device's abort timer. #799
  • A staged IOS XE image that passed a check is not checked again while its index, size and modification time in dir stay the same. #799
  • The prepare job tells how far a transfer is, in progress and in its message: over NETCONF from the size of the file on bootflash, over the CLI from the output of the copy. A software adapter reports it with the new SoftwareManager.report_transfer. An xcopy also ends when the file has the image size, as IOS XE 17.9.5a records no xcopy. #799
  • No upgrade runs while the device is not connected: the software manager waits as under a veto and reads the release again when an adapter connects. #799
  • The install is a job in the device software state, job[install]. While the prepare or the install runs, the message of its job says what the device does, as the adapter reports it with the new SoftwareManager.report_step. After the install, the install job says when the commit window ends, until the run commits or rolls back. The IOS XE adapter reports the transaction that the device runs, like add: extract or activate: reload, the step lines of an install command over the CLI, the check and the transfer of the image, and waiting for the device to boot. The IOS XE CLI adapter has install-progress, the last step line of the install command that runs. The IOS XE mock lists a running operation in Cisco-IOS-XE-install-oper and takes it through its stages. #803
  • The software intent has an attempt, 0 by default. The state has an entry for each attempt at the target in the campaign, state/attempt[number], with its status, stage, reason, checks and jobs; the entries of earlier attempts stay. A final status stands for its attempt: a reconnect, a new device type, or the target removed and set again runs nothing. A new attempt reads the device again, and runs a failed upgrade again when no veto holds the install. A run of another attempt that is under way finishes first. When that run fails its rollback, the device may run the release it installed, uncommitted, so the new attempt does not start: it answers the new status not-started, with the failed attempt in its reason. A run that a new target or campaign, or a removed target, stops before its install ends at upgrade-needed. #804
  • Images can be kept in folders inside the image folder. An image in /images/image has a file leaf, the path of its file there, such as iosxe/17.18/c8000v-universalk9.17.18.03a.SPA.bin. Without it, the file is the image's name. A device's URL is url-base followed by the file. Campaigns keep naming the image, so moving its file changes only the file leaf, and the campaigns give their members the new URL. Members that answered or finished keep their status, since the target release stays the same. GET /image-file lists the files in every folder, by path, and the folders under folders. PUT /image-file/<path> makes the folders on the way. PUT and DELETE /image-folder/<path> make a folder and remove one with everything in it, and POST /image-move with {"from": ..., "to": ...} moves or renames a file or folder without replacing anything. #806
  • The software state shows the vetoes of the configuration under state/veto, with their reasons. #811

Fleet Manager

  • Campaigns show each device's pre-check and post-check verdicts, upgrade stage and failure reason, and the web UI draws each run as a chain of steps. How long the checks wait is set per device under upgrade-checks. #667 #678
  • Fixed replanning a campaign moving done and in-flight devices to the time of the replan. Released devices keep their real times, and only the waiting ones are placed again. A window that is already open is planned with the time it has left, instead of counting the devices that don't fit as late. #672
  • Campaigns are paced by how many devices upgrade at the same time, with target-concurrency and max-concurrency, instead of target-rate and max-rate in devices per hour. With rates, slower upgrades put more devices in flight, which overloaded the system further. #723
  • Fixed campaigns that never ended after a window of a schedule opened again while its earlier window still held released devices. The waiting devices were placed in both windows and counted twice. #722
  • The release decisions of a campaign moved from the campaign actor into the stratoweave.fleet_ctl.Controller class. The controller has no clock and no I/O, so test_fleet_ctl runs it without an actor. #725
  • Removed the stratoweave.fleet_sched module. Its helpers moved into stratoweave.fleet_ctl, and the campaign simulator stratoweave.fleet_sched_sim is now stratoweave.fleet_ctl_sim. #740
  • A campaign's plan comes from the controller that releases the devices, run ahead at the expected upgrade durations. Before, a campaign could plan without alarms and raise alarms as soon as it ran, especially in fleets that mix fast and slow devices. Each device keeps the window its plan gives it, and a device that misses its window moves to the next window of its schedule. The controller goes above target-concurrency, up to max-concurrency, only when the windows or the deadline need it. max-concurrency counts the devices of the whole campaign. Plan alarms name the affected devices. A window that opened on an earlier day and is still open is used too. #730
  • Devices can be members of redundancy groups, set under /fleet/device/redundancy-group with a redundancy-priority. A campaign upgrades the members of a group one at a time, lower priority first and equal priorities by name. /redundancy-groups sets how far apart they go: spread, the default, moves later members to later windows of the campaign where a device of the same estimate in no group can trade windows with them, and pack starts each member as soon as the one before it is done; both keep an optional min-separation. When the upgrade of a member fails or rolls back, the other members keep their windows and wait as the group's failure-policy says. The plan's alarms name the devices that wait. #740
  • Fixed upgrade duration estimates being reduced by devices that already run the target release. These devices could cause the campaign to underestimate upgrade times and schedule too many devices in the first windows. #740
  • A campaign no longer raises its concurrency for good when a device reports a few seconds late or upgrade times scatter around the estimate. The expected pace is the mean of the last upgrades, a round and a quarter at target-concurrency, and a device is late only a fifth of its upgrade after it was expected to end. Before, one late device could raise the concurrency for the rest of the campaign and give alarms. #746
  • A window starts a device only while its close margin is left before the window closes, by default one and a half times the device's slow time. The slow time is the 98th percentile of the last upgrade times, two rounds at max-concurrency and at least 50, so the margin grows when slow upgrades show, and one slow upgrade alone moves it only through the mean of the last upgrades. Set the margin with close-margin on the campaign. Before, a fifth of the expected duration was the only margin, and upgrades in a heavy tail ran past the close. A device whose margin is longer than the time left in every window of its schedule does not start, and the plan says so. #746
  • The ramp lets three more devices start for each device that is done without a failure, up to the concurrency, instead of doubling once a whole round has finished. A round now takes four times as many devices, the starts spread out as devices finish, and when the plan raises the concurrency, the extra devices start as devices finish instead of all at once. Failed upgrades no longer grow the ramp. #746
  • A campaign spreads its starts over the stagger, in percent of the expected upgrade time: 25 by default. A device that starts holds the next start back by that part of its upgrade time, divided by the concurrency, so a larger campaign starts faster. Before, a window that opened with room for many devices started them all in the same second. #748
  • Campaign plans are computed faster. The first plan for 100,000 devices takes seconds instead of minutes, and a decision that only starts devices reuses the last plan while no report came in, no device is due and no window opens or closes. #748
  • Campaign decisions are about 6 times faster at 100,000 devices: a plan after a report takes 85 ms instead of 521 ms in a release build. Most of the gain is in the run ahead that computes a plan. The controller numbers the devices and keeps their state in lists by number, so a run ahead copies and indexes lists instead of hashing device names in tables of 100,000 entries. Each step of the controller takes the reports that it has not counted yet from a heap ordered by time, instead of going through all of them. The windows of a schedule share one count of their waiting devices and one sort of the devices by estimate, instead of one per window. The campaign reads its members' schedules, estimates and redundancy groups from the inventory when the campaign or the inventory changes, instead of at every decision. #787
  • Fixed a member whose final report came before its release, but in the same second, counting as an upgrade of one second. It happened to members that were already up to date when the campaign started, and dragged the campaign's pace down. #766
  • Maintenance windows that are open at the same time share a campaign's free slots. Before, the window that opened first took every slot it could use, and a window that opened later waited until the first had nothing left to start. Each schedule with an open window now gets slots by the upgrade time it must start before the first of their last starts, counting the time it has after the others close. So a schedule with time of its own at the end leaves the slots to one that closes first, and a deadline needs about the concurrency it needed before. Windows that close together share in proportion to the upgrade time waiting in them. #770
  • Fixed an edit of a campaign, or of its schedules, images, redundancy groups or member devices, removing the software intent of the devices that the campaign had released, until the campaign wrote it again a moment later. An upgrade that had not reached its install could fail. #783
  • Campaigns acquire their members in plan and in run. Each member in the fleet gets the campaign's name, its target release and image URL, and a veto with the holder campaign:NAME, which holds the install. So each member reads the release it runs and answers up-to-date or upgrade-needed, and nothing is installed. A member added to a running campaign is acquired the same way. In run, the campaign removes a member's veto when it releases the member, and puts it back when the member is done. Before, a campaign wrote nothing until it released a member in a window. #783
  • A device in the fleet is in one campaign at most, also while the campaign is in plan and after it is done. An edit that names a device in a second campaign fails with "Cannot merge leaves with different values at /software/upgrade-campaign". To move a device, remove it from its campaign, or delete the campaign, in an earlier edit. A device that is not in the fleet gets nothing from a campaign, so two campaigns can name it. The edit that adds it to the fleet then fails with the same error until only one campaign names it. #783
  • A campaign's device-status shows a member's status only when the status is for the campaign's intent, its name and the member's target release, and pending otherwise. The counters count only those statuses. Before, a succeeded left from an earlier campaign made a member done at its release, with no upgrade. #783
  • Setting a running campaign back to plan pauses it: members in flight stop before their install, and are released again when the campaign runs again. A member that still reports in-progress for the campaign when it runs again, or after a restart of fleetmgr, counts in flight and keeps the campaign's veto until it is done, or until it stops before its install. Stopping a campaign and running it again no longer retries its failed members. #783
  • With allow-staging, a campaign in run lets each member that needs the upgrade, and that the plan gives a window, prepare its image once before its window, under the campaign's veto. The first member prepares at once. The others follow as far apart as their deadlines allow: the opening of the next window of the member's schedule, less close-margin percent of its expected prepare time. No member prepares while a window of its schedule is open, and a failed prepare is not tried again. In plan nothing is copied. The plan's new prepare-start says when each member prepares, and prepare-concurrency how many copies run at most at once; there is no cap. device-status shows each member's prepare, and the new counters preparing and prepared count the members that wait for their release. New plan alarms: "N device(s) cannot prepare before their window opens" and "N device(s) failed to prepare and copy the image in their window". The upgrade in the window still copies the image again. Before, allow-staging on a campaign had no effect. #784
  • A campaign releases a member only once the member's status answers the campaign's target: upgrade-needed, or a final status. Until then the plan holds the member in its window, with the alarm "N device(s) have not answered the campaign's target", and the members after it in its redundancy groups with "N device(s) wait behind a redundancy group member that has not answered". unknown and pending are no answer. A campaign in plan plans again as its members answer, and plans a run that starts now, so its plan shows what run would do. So a member added to a running campaign waits until it has read the release it runs, and is then planned and released. A campaign created directly in run promises its windows in the order its members answer. Create it in plan first to keep the campaign's order. A decision that only starts devices still takes the last plan again when the plan's only alarms are for held devices. #783
  • A campaign's new failure-budget, in percent of its members, stops its releases when more members have failed than the budget. The members in flight finish, and the plan's alarm says "N device(s) failed, more than the failure budget of M: no more devices start". To go on, raise or remove the budget, or remove failed members. Without a budget, a campaign releases as before. #791
  • A campaign holds a waiting member while the member is down: its flotilla is not connected to it, or the top is not connected to its flotilla. The member starts when it is back up, so a dead flotilla holds only its own members. Before, a member that was down was released and failed. Each held member's device-status gives the reason in held and the device or flotilla that holds it in held-by, and state/held counts them. New plan alarms name the members that are down, the members in flight that are still down after their expected end, and the held members that lost the window promised to them. #794
  • A campaign's new notification-period, in days, gives the customers of its members time to be told about their window: the campaign uses only the windows that open that long or more after it first went to run. That time is the campaign's state/run-since, and a pause and a restart of fleetmgr keep it. For a campaign from an earlier version, run-since is the time Fleet Manager first starts on this version, or, for a campaign in plan, the time it next goes to run. 0, the default, gives no notice. #795
  • A campaign member has an attempt. Changing it tries the member again: the campaign takes it as a new member, its device reads its release again, and a member that failed is released again in a window. A member in flight gets the campaign's veto back; a run past its install finishes and holds its slot until it ends. Removing a member and adding it again no longer tries it again: its device keeps the final status of its attempt. device-status shows the attempt that counts, the one in progress or else the member's attempt, in its new attempt, and no longer the upgrade-campaign and target-release of the device. A member with a new attempt no longer counts against the failure budget. A new attempt set while a run fails its rollback does not start: the member shows not-started, counts as failed, and holds its redundancy groups as a failed rollback does, until it has another attempt. #804 #814
  • A redundancy group's new failure-policy says what a failed member does to the other members. With continue, the default, a member that failed or rolled back holds them only until it is done and up again, and the next member starts after the group's min-separation. A member whose rollback failed is in an unknown state, and holds them until an operator acts. With hold, the other members wait for an operator after any failure, as before. #811
  • A campaign holds a member whose device has a veto of another holder, such as a service health check, until the veto goes. device-status shows the device's vetoes, and the held reason vetoed with the veto's holder in held-by; the members after it in its redundancy groups are held as group-vetoed. The plan's alarm says "N device(s) are held by a veto of another holder". A released member that gets such a veto before its run starts, or that the veto stops before its install, waits again and frees its slot. Before, the campaign released it, and it held a slot while its device did not install. #811
  • Each /fleet/device says in state/connected whether it can be reached: connected, device-down while its flotilla is not connected to it, or flotilla-down while the top is not connected to its flotilla. A device that was down before the top lost its flotilla stays device-down, as a campaign holds it for down. state/connected is absent while the device has not connected since fleetmgr started. A campaign's new state/up-to-date counts the members that answered up-to-date, as they already ran the target release when they read it; succeeded still counts them too. #816
  • A campaign member whose status is final while it waits is done at once, without a release: a member that is up to date, or one that succeeded or failed before a pause. It takes no window and no slot, and adds nothing to the pace or the ramp. The next member of its redundancy groups waits the group's min-separation from the answer of a member that is up to date, or, when that member upgraded in the run before it got a new attempt or joined again, from the end of that upgrade. After one that upgraded before a pause it waits the min-separation from that member's report. A failed one still counts against the failure budget and holds its groups as their failure-policy says. The plan shows it in no window, and the campaign page counts it as already done. It waits again when its final status goes. Before, it waited for a window of its schedule and held a slot there until the next decision. #830
  • Each /fleet/device has the platform that its flotilla reads from the device under state/platform, absent until the device has connected. A campaign's device-status has each member's version and model instead of running-release. The web UI shows the platform on the devices, device and campaign pages, and the devices page filters by model. #718
  • The top always starts ten flotillas and places each device on flotilla-N by a hash of its name. Removed /fleet/node and /fleet/device/shard. Devices without a shard, which the top managed itself, are no longer supported. #713
  • Campaigns select an image for each device instead of setting one target-release and image-url for the whole campaign. Images must be listed under /images/image with a version. The campaign does not configure an upgrade for a device if its image is missing from the list or has no version. An alarm identifies the image and how many devices are affected. The campaigns page no longer has a Release column. #743 #759
  • The new campaign page has three steps: pick devices by filtering or from pasted names, pick an image for each model in the selection, and create the campaign, which opens its page to set the schedule and pace. A family's default image is preselected when exactly one default image lists the model. Devices that haven't reported a model are left out, and a model that no image lists prevents campaign creation. #743
  • The web UI has image pages. Image files dropped on them are uploaded, with OS, version and family filled in from IOS XE file names. An image can be made the default of its family, and the images page sets the image list's url-base. Deleting an image also deletes its file. #709 #743
  • The web UI's images page browses the image folder like a file manager: open folders, make them, upload into the one shown, and rename, move and delete files and folders, from a row's menu, by dragging rows onto a folder, or on several selected rows at once. A move changes the file of the images it takes along and keeps their names, also for images that campaigns install. An upload whose file name is already an image's name fails before the file is sent. Search covers the folder shown and the folders in it, and files that the image list lacks show, ready to add. #806
  • The web UI's image pages have a Checksum field for the checksums that the vendor publishes: MD5, SHA-256 or SHA-512, told apart by their length. An upload sets the image's size from the file. #823
  • The web UI has a Config page that shows the top's configuration, the state its transforms publish, or both, as XML, JSON or AData. The Apply config page merges JSON or XML configuration into the running system with a RESTCONF PATCH, after a confirmation. #736
  • The web UI has a search, in the header or with Ctrl/⌘+K, that jumps to a page, campaign, image, device or schedule by name. #737
  • The web UI has redundancy group pages: a list, one page per group with its members in upgrade order and its separation, and a page to create a group. A member's priority can be typed there, or the member dragged to another place, which sets the priorities that give that order. A change that reorders another group of the device asks first. The device page sets a device's groups and priority, and the devices page shows and filters by group. The demo fleet pairs devices in groups. A campaign's page shows the groups of its members, each member in upgrade order with its status and time. Groups held behind a failed member come first, with a button that takes that member out of the group. #789
  • The campaign page has Schedule and pace under the plan, folded each time the page opens. The failed devices, the redundancy groups and the devices fold too, and only the failed devices start open. Folded, the redundancy groups say how many of them are held. #789
  • The campaign page's Revert to plan is now Pause, and its confirmation says what a pause does: members that upgrade stop before their install and go on when the campaign runs again. A campaign in plan that has run shows as paused. Schedule, pace and limits set the campaign's close-margin and stagger. #823
  • The device page is redesigned around the device's upgrade run. #674
  • The web UI's add and import device forms take a username and password, instead of writing admin/admin for every device. Import uses one pair for all its devices. #726
  • The web UI's add and import device forms take a private key file, and a CLI over SSH box that gives the device the iosxe-cli extra adapter, on port 22 or the port typed. A device with an address needs a password or a key. Import gives the CLI only to the devices with an address. #823
  • The web UI works on narrow screens and has better contrast and accessibility. The campaign page leads with live progress and failures. #679
  • The web UI reads less from the server. Each page reads only the fields that it shows, with the RESTCONF fields query parameter, and the device page no longer reads the device's credentials. The pages that read the whole fleet, like the devices page, share one read for 30 s, so a change that another client makes can take that long to show. #671 #689 #807
  • Web UI pages show at once and fill in as their reads arrive, and the sidebar marks the page that is loading. On large fleets the pages read less: the device page reads only the device's own device-status entry, a large campaign's page reads its state less often the longer the read takes, and adding, importing, removing or scheduling devices updates the devices list without reading the whole fleet again. Newly added devices show no model or version until the page is loaded again. #754
  • Fixed the campaign plan timeline drawing a crowded window past its end. Its columns grow taller instead. #761
  • The campaign page shows why members wait. Its progress bar splits the members that it upgraded from those that were up to date already, and the members that wait into held, not answered and waiting, and says what holds the held ones: the device or its flotilla being down, a veto of another holder, a member of its redundancy groups, or the failure budget. Before, up-to-date members counted as succeeded, and every waiting member as pending or unknown. The devices table says why each held member waits, links the group member that holds it and filters by the reason. The plan timeline draws members that are down grey and the other held members amber, with a legend of the colours it shows, and leaves out the members that already run their target, which are listed above it. The redundancy groups take whom each member waits behind from the campaign, and offer to take out only the failed members that others wait behind. A pending member shows as not answered, and the device page names the attempt it has not answered. #816
  • The campaign page sets the campaign's notification period and failure budget, shows how many failures the budget allows and how many there are, and when the campaign first ran. #816
  • The new campaign page shows, for each model, how many of the selected devices already run the picked image and how many run a newer release. It leaves out the devices that run their image, unless told not to, and warns that the campaign would downgrade the newer ones. #816
  • The devices and device pages show whether each device can be reached. #816
  • The redundancy group pages set a group's failure-policy next to its separation, and the groups list shows it. On a campaign's page, each group gives its policies under its name, a member whose rollback failed says so, as it holds its groups whatever their policy, and the confirmation to take a member out says why the group waits for it. #819
  • fleetmgr and flotilla run Acton's generational garbage collector. Start them with ACTON_GC_MODE=ordinary for the ordinary collector. #764
  • Removed the unused cpe-device model. The IOS XE device type directory is now dev-iosxe. #694
  • just demo-all deploys the CEs, starts the top and loads both campaigns. #690
  • just run starts the top and its ten flotillas with no devices, images or campaigns. #736
  • just build builds fleetmgr and flotilla with --release, optimized for speed and without debug symbols. #805
  • The web UI is built with SvelteKit 3, and just gen-webui needs Node 22.17 or newer. #739
  • A fleet device can have extra adapters, as a StratoWeave device can: the extra-adapter list with the type, address, port, credentials and SSH settings of each, which the transforms pass to the device on its flotilla. The flotillas register the device type iosxe-cli, the IOS XE CLI adapter, so an IOS XE device can have the CLI beside NETCONF. #779
  • A fleet device has upgrade-checks/enabled, true by default. With false, its upgrades go without the pre- and post-checks. #799
  • The flotilla checks a device's interfaces and OSPF neighbors before and after an upgrade, as it checks BGP. The interfaces post-check wants every interface that was up, line protocol up, before the upgrade to be up again, NVI and Virtual-Access interfaces aside, and as many Virtual-Access interfaces up. The ospf post-check wants the same OSPF neighbors, each in the same state. Their pre-checks only record what they read. The checks are registered as bgp, interfaces and ospf instead of rfs-device-health, so a verdict says what each check saw. As before, the device is read only while a check runs, and upgrade-checks sets the waits of all three or turns them off. No adapter reads the interfaces or the OSPF neighbors of an IOS XE device without NETCONF, such as 16.8 on a C1111, so its pre-checks refuse unless the checks are off. #821
  • The prepare in the upgrade status of a fleet device, and in the device-status of a campaign, has progress: how far the device's transfer is. #799
  • The upgrade status of a fleet device, and the device-status of a campaign, have install: the device's install job, whose message says what the device does. The web UI shows it for each device of a campaign and on the device page, with a bar for the transfer of the prepare. #803
  • just demo generates its fleet like just demo-fleet: a hundred devices of five models, most on 17.15.04, some older and a few already on 17.18.03a, and a campaign over 80 of them whose windows open within three minutes. The other 20 are left to pick for a new campaign. The demo images are four families in three releases, each family in its own folder, with an empty sparse file for each in out/images. demo.xml, demo-campaign.xml, upgrade.xml, just upgrade and just upgrade-and-watch are gone. #825
  • A /fleet/device has the device's last-boot in state/platform. [#841]
  • A member that is done without a release, and whose upgrade the campaign did not see end, is done when its device last booted, when the device tells. After a restart of fleetmgr, the next member of its redundancy groups then need not wait for a device that has run the target for months, and waits the min-separation from the boot of one that booted a minute ago. When the campaign saw the member's upgrade end, in the run or before a pause, the member is done at that end, or at a later boot. A member whose device does not tell is done at its answer, as before. [#842]

Dependencies

Acton

  • Requires Acton newer than 0.31.0, for explicit parameter defaults acton#3211, the Env services acton#2768, the write barriers of the generational garbage collector acton#3390, HTTP server bodies as bytes acton#3403, compressed HTTP responses acton#3404 and the raw terminal mode of stdin acton#3419.

acton-yang

  • gdata is encoded as compact XML and JSON by default, as in the Acton stdlib. Pretty output was only meant for developers. acton-yang#541 #684
  • Removed the YANG parser statement cache and the yang.compile_with_cache() function. With the recent improvements to Acton RTS and GC that part of the pipeline is no longer significant. ~/.cache/ayang is not used any more. acton-yang#544
  • DRoot.module_metadata entries are yang.schema.ModuleMetadata tuples, with the module in name instead of module. The new submodules and implemented fields are for schema discovery APIs. acton-yang#545
  • Optional APIs and generated constructors declare explicit defaults.
  • Fixed YANG patterns with a . or $ in a character class. The nokia-conf module of SR OS 25.10 failed to compile with "Invalid range bound", and [.] and [$] matched any character and the end of the string instead of a literal dot and dollar sign. acton-yang#546 #729
  • YANG sources are parsed concurrently, which makes compiling large module sets faster. Removed yang.parse_modules() and yang.compile_from_stmt(), which were only needed for the statement cache. acton-yang#550
  • yang.schema.DRoot.merge() merges compiled schema roots into one, like the CFS schema and ietf-yang-library that a northbound server serves together. It raises ValueError when two roots hold the same top-level node or different revisions of a module. acton-yang#548
  • take_leaf() and take_leaflist() of the yang.data.YangData protocol return no t type field. gdata keeps a union's value as its Acton type, so the field is no longer needed. acton-yang#553
  • The take_container(), take_list(), take_leaf() and take_leaflist() methods of yang.data.YangData no longer take the node's adata attribute name. The adata generator stores that name in the schema as the new adata_name of each node, and yang.data.from_data() reads it from there instead of computing it on every call. acton-yang#552
  • yang.schema.DNode.gname is the node's gdata.Id instead of the name of its gdata class, like Container, and a DList has the gdata.Id of each key in key_gnames. yang.data.from_data() uses them instead of creating a gdata.Id for every node it reads, and the generated adata from_gdata() methods use gdata.Id constants of their module. Fleet Manager with 50,000 devices has all of them connected in 22 s instead of 33 s. acton-yang#557 #762
  • Fixed subscription data callbacks running in the provider's actor. yang.gdata.Dst and the data callback of TreeProvider.declare_subscriptions take an action again instead of a proc, so a consumer that passes its actor method gets the data in its own actor. A generated dst() converts each tree in an actor of its own instead of in the provider. acton-yang#556 #758
  • Fixed XML parser skipping a leaf with an empty tag, like <k1/> or <k1></k1>. Its normalized value is an empty string. This makes it possible to use the empty string as a key value too. acton-yang#554 #744
  • A root_path passed to yang.data.from_data() can lead to a notification, top-level or nested in a container, so a notification payload decodes like RPC output. acton-yang#555
  • Fixed XML decoding keeping only the first of a container element that occurs more than once in the same parent. IOS XE 17.15 periodic yang-push updates repeat <neighbors> for each BGP neighbor, so a NETCONF subscription got only the first neighbor and the fleetmgr BGP post-check failed. The occurrences are now read as one container. acton-yang#560 #802
  • yang.restconf.resolve_fields() parses a RESTCONF fields value against the schema below the target resource, and select_fields() narrows data to the result. acton-yang#561
  • yang.compile_with_sources() compiles like yang.compile() and also gives each source with the name and revision of its module or submodule, as a yang.schema.YangSource. DRoot.prschema() and DRoot.prdaclass() take their schema_yang sources in that form, and the generated src_yang() functions return them. acton-yang#562
  • Fixed subtree filters dropping a list entry when one of its selection nodes selected nothing. A fleet device in no campaign has no software state on its flotilla, so the top's subscription got nothing and the device showed no platform or connection. A content match node that matches is now in the output too. acton-yang#563
  • Imports Acton library modules from std, and the schema code generator writes import std.base64 and import std.json into the modules it generates. acton-yang#565
  • Fixed the compile failing on an augment whose target is inside a structure: tailf:structure, sx:structure or rc:yang-data. A structure describes a message, not datastore data, and its contents are not compiled, so the augment is now skipped. An augment into any other extension still fails. ConfD's tailf-rollback augments a structure of tailf-yang-patch like this. acton-yang#564
  • Fixed relative leafref paths in an action, an rpc or a notification. The action, rpc or notification is the parent of its leaves (RFC 7950 6.4.1), but the path left it out, so .. went one level too far, and in an rpc ../a resolved to a top-level leaf a. A path into another action or rpc than the one that holds the leafref now fails. acton-yang#564
  • A leafref path can have whitespace between two predicates, like acl-set[name=current()/../name] [type=current()/../type], as XPath and ConfD allow. The path used to end at the whitespace, so it pointed at the wrong node. A path with any other trailing input now fails to compile. acton-yang#564
  • A JSON leaf with the value null fails to decode, as RFC 7951 requires. acton-yang#566

acton-ssh

  • Fixed SSH events being delivered out of the order libssh raises them. Under load, pipe streams were reordered and broke NETCONF chunk framing. acton-ssh#15
  • Fixed a Client hanging forever when its on_hostkey handler never responds. connect_timeout also covers the on_hostkey decision. acton-ssh#6
  • SshPipe takes connect_timeout and auth_timeout. Before, it always used the Client defaults of 10 seconds. acton-ssh#18
  • SshPipe can request an interactive shell with shell=True, instead of a subsystem or a command. acton-ssh#20
  • Builds with Zig 0.17. acton-ssh#17
  • Imports Acton library modules from std. #833
  • Connect failures are reported as ssh.SshError, or its subclasses KeyExchangeError, HostKeyError and AuthenticationError, instead of an OSError for everything. An OSError is left for failures before the server's SSH banner, like a refused TCP connect. Client.on_connect gets an exception instead of a string. acton-ssh#21
  • Stores GC pointers with write barriers, so programs that link it can run in Acton's generational GC mode. acton-ssh#23
  • ssh.Server serves interactive shells when it gets an on_shell action, which receives the client's terminal as a PtyRequest, or None when the client asked for none. Without on_shell the server refuses shells, as before. Channel.change_pty_size() sends a terminal size change, which a ServerChannel receives with on_window_change. acton-ssh#24
  • Fixed writes that queue more than 8 MiB failing their channel, which kept a NETCONF message over 8 MiB from going out over an SshPipe. max_write_buffer of Client and Server defaults to 0, no limit. Writes go to libssh in pieces of at most 256 KiB, which also bounds libssh's send buffer per connection. acton-ssh#25
  • ssh.ScpPut and ssh.ScpGet copy one file to or from the server over SCP on an existing Client, without blocking. They run the server's scp on an exec channel and hold the whole file in memory. acton-ssh#22
  • Fixed a busy loop while a channel has writes queued and the peer's window is shut. acton-ssh#22
  • A failed connect or a dropped connection reports the socket's error, such as "connection refused" or "no route to host", instead of "bad file descriptor". acton-ssh#22

acton-lmdb

actmf

  • Updated for explicit defaults, for Acton library modules imported from std, and for the changes to the yang.data.YangData protocol and yang.schema.DNode.gname in acton-yang. actmf#76 actmf#78 actmf#79 #741 #790

actsnmp

  • New dependency: an SNMP implementation in Acton, with SNMPv1, v2c and v3 and a MIB compiler that emits YANG (RFC 6643). The SNMP device adapters use it. #701

http-router

  • Query parameter names and values are percent-decoded. Context.query() gives every value of each name as a list, in request order, instead of one value, and query_param() gives the first. #805
  • Imports Acton library modules from std. #833
  • The respond callback of handlers and middleware takes the body as bytes instead of str, as http.Server in Acton does. acton#3403

Documentation

  • The software upgrade reference is rewritten to match the code, and is part of the guide. #697
  • The guide has a Releases page, which includes this changelog, and a Roadmap page. #800

Testing & CI

  • Downstream upgrade PRs run make gen, so changes to generated code don't break the downstream's checks. #706
  • Removed the actor pipe restart stress test, which was too slow for CI. #696
  • Fixed the build of the multi-arch ghcr.io/stratoweave/ncurl:latest image on runners with Docker Engine 29. #735
  • The notconf tests connect to 127.0.0.1 instead of localhost, which macOS resolves to IPv6 first, where Docker's published port does not answer. #751
  • Fixed just system-test in fleetmgr timing out. It runs each system test once. #747
  • The gc-check CI job runs the fleetmgr tests with acton test --gc-check and the fleetmgr demo for a minute under the GC write-barrier checker. #764

Compatibility Notes

  • Requires Acton newer than 0.31.0. Regenerate generated code with make gen after upgrading. #704 #764
  • adapter_type and software_adapter_type in a DeviceType take ?Env instead of ?net.TCPConnectCap. Adapters in other repositories, like sw-adapter-cisco-ios-cli, need the same change. #702
  • gdata to_xmlstr() and to_yang_jsonstr() return compact output by default. Pass pretty=True for indented output. acton-yang#541
  • /restconf-state/capabilities no longer lists YANG modules. Read them from /ietf-yang-library:modules-state or /ietf-yang-library:yang-library instead. #676
  • The NETCONF server's <hello> lists no module capabilities, only yang-library:1.1. Clients that take the server's modules from the <hello> must read them from /yang-library or /modules-state. #716
  • The device software state has no running-release. Read the release a device runs from state/platform/version, and in Fleet Manager from the version of a campaign's device-status. Mock devices take the release they start on as mock/platform/version instead of mock/software/running-release. #718
  • An app with its own device_types.act must add import stratoweave.platform_probes as swplat and platform_probes: list[swdev.PlatformProbe] = swplat.PROBES to it. The generated sysspec.act reads platform_probes from it, and the generator does not change an existing device_types.act. #712
  • Fleet Manager has no /fleet/node and /fleet/device/shard. Remove them from its configuration. #713
  • Fleet Manager campaigns take target-concurrency and max-concurrency instead of target-rate and max-rate. Move configuration that sets the old leaves. A campaign that sets neither upgrades 10 devices at once and goes up to 50 when its windows or deadline need it. There is no unlimited value. #723
  • Fleet Manager campaigns take an image in each device entry instead of target-release and image-url. Add the images to /images/image, with their version, set /images/url-base, and name an image for each member. #743
  • A device can be in one Fleet Manager campaign only. Before you upgrade, remove each device that two campaigns name from one of them. Otherwise Fleet Manager cannot load its configuration and does not start. #783
  • Since a NETCONF <edit-config> transaction completes before the devices have applied their configuration, a device rejecting its configuration is no longer reported as an rpc-error. The device side is visible per device instead: state/adapter shows whether it is connected, and the inspection API's /device/<name>/info and /device/<name>/log show its queue and what was sent to it. #792
  • DeviceMgr.get_config() takes a done action and gives the running configuration that the DeviceMgr keeps, or reads it from the device. It returned the adapter's copy before. get_running_conf() gives None until the first read.
  • stratoweave.main() takes static_assets as dict[str, bytes]. Have the web UI build embed each file as a bytes literal, as fleetmgr's gen-embedded-assets.mjs does. http-router handlers and middleware call respond with a bytes body: call .encode() on a str body.

26.10 - 2026-10-23

Hello, world! This is the first formal release of StratoWeave, woven from first principles.

There’s still work ahead before it’s ready for production, but it’s coming together quickly, and we think it’s ready for people to pick up and try. This release is an invitation to explore, experiment, and help shape what comes next. Give it a spin. We’d love to hear what you think.