Yeah being robust to dropping a context update is a massive pain :( The trace interning perfetto.dev/docs/instrum... has this issue, you have to be able to detect both drops and ring buffer waps and re-emit the interned data.

The improvement over naive compression can be significant (e.g. just delta encoding sched switch was a larger improvement than gziping the whole trace) to the extent we felt compelled to rush it in for Android P/Q (I forget which). On a macrolevel a trace is 100% self contained so still suffers!

On a microlevel Perfetto does this (delta encoding) in some cases. The two big ones that come to mind are sched switch and process metadata. Process metadata = big initial dump + delta updates. Delta encoding timestamps in sched takes advantage of varint in protobuf (delta ns is 2 bytes, abs is 8).