node-red-contrib-ccu 4.8.0

v4.8.0

on Node.js NPM

Less memory: the connection's own bookkeeping no longer grows with the
number of datapoints, and the paramset descriptions in memory are those of
the device types the system has. On a large system (a few thousand
datapoints, a few hundred subscriptions from value nodes, RedMatic-HomeKit
and RedMatic-Matter) that is on the order of 100 MB of live heap less, and
a deploy no longer adds to it. It is a reduction, not the whole of a large
Node-RED's memory: flows and the other nodes keep what they keep.

Changed

• Datapoint subscriptions are dispatched through an index. Until now the
  connection remembered, for every datapoint, which subscriptions had matched
  it and which had not - two sets of ids per datapoint name, so the memory
  grew with datapoints × subscriptions (3,500 datapoints × 536 subscriptions
  were about 70 MB of live heap on a CCU3), a subscription with cache on
  walked every cached value at deploy time to fill them all, and - because
  the replay and the event path stored the id in two different forms - a
  redeploy left the old entries behind and grew the sets again. Now a
  subscription is indexed by the most exact attribute of its filter (the
  datapoint name, the channel, the device, ...), an event is tested only
  against the subscriptions that can match it, and unsubscribe removes
  exactly one entry. What a subscription receives, and in which order, is
  unchanged (a randomised equivalence test against the old dispatch guards
  it). In the benchmark (test/bench/subscriptions.js, 12,000 datapoints,
  1,000 subscriptions) the heap goes from 481 MB to 3 MB, a deploy from
  6.5 s to 40 ms, an event from 28 µs to 8 µs.
• Paramset descriptions in memory only for the device types present.
  paramsets.json - the shipped 14 MB plus what a system fetched - was held
  in memory whole: some 2,800 type/version keys, 16-20 MB of heap, where a
  system uses the few dozen types it has (40 types are 3 MB). The file on
  disk keeps every description in its format; the memory holds the ones the
  device table needs, loaded at start and, for a device that arrives later,
  from the file before the interface is asked. A fetched description is
  merged into the file. The editor's device data is correspondingly smaller.
• An obsolete ccu_paramsets_v2.json (the format of 2019) still lying in a
  userDir is removed at start, once, with a log line.

Commits since v4.7.2

• 4.8.0 - less memory: subscriptions through an index, paramsets of the present types (7ec00d4)
• docs(changelog): 4.8.0 (102462c)
• test(lite): wait for the listDevices call instead of asserting it at once (119ccdb)
• feat(connection): paramset descriptions in memory for the device types present (c44d726)
• feat(connection): dispatch datapoint subscriptions through an index (ecc8a72)
• test(paramsetfault): wait for the interface's device table before the queue check (a12883f)