The old bug in macOS Sierra, which has been present since its first release last year, resulting in Time Machine backups and other background tasks becoming highly irregular and unreliable, has not been fixed in Sierra 10.12.4.
This bug does not affect Macs which are shut down daily, or even weekly, for example. It also may not affect Macs which are left running continuously for many days, but which enter system sleep; I am unable to let my Mac enter system sleep because when it does so, it sleeps its hard drive (despite being set not to).
After a variable period of continuous running, usually longer than 7 days, sometimes over 30 days, Time Machine backups and other activities which are managed and dispatched by the Duet Activity Scheduler (DAS) will cease being scheduled correctly. They then become extremely irregular and unreliable – backups may be made several hours apart, for example. All activities managed by DAS are affected.
I reported this bug to Apple in February, and initial experience with Sierra 10.12.4 suggested that it may have been fixed. It has not, but persists in 10.12.4.
Those systems most likely to be affected are those running macOS Server, as reported in comments on this blog.
The only solution is restarting the Mac when backups become irregular.
DAS is part of an undocumented system in macOS which I have previously referred to here as Grand Central Dispatch (GCD), which is incompletely documented – and have been criticised in comments for doing so. My DispatchView tool is an excellent way of diagnosing this problem, and I have the log extracts to demonstrate how this bug occurs. Still. Whatever anyone likes to call the affected system.