Partition timeseries with state using Dask Dataframe based on custom condition - dask

Question

I have a large number of parquet files in s3 which together represent a large timeseries describing an evolution of a certain state
delta timestamp
0 True 1
1 False 2
2 False 2
3 True 3
4 True 4
5 True 5
6 True 6
7 False 7
8 False 7
9 False 7
10 True 8
11 True 9
Rows grouped by same timestamp with delta==False represent a full state of a system at given timestamp (note same timestamp values), while delta==True represent a partial change of the previous state.
My goal is to apply a row wise operation to this timeseries which restores state, updates it on each change and performs certain calculation, producing a new timeseries (or an aggregate value, this is not relevant at this point). All of this should be performed in parallel on a multi-machine cluster with Dask.
For this, I want to create a partitioned Dask Dataframe which makes sure that each partition starts with a block of rows with delta == False e.g. each partition knows full state at the beginning and apply my function to this partition.
Partition 1 (note row 0 is discarded since no previous `delta==False` values are found)
1 False 2
2 False 2
3 True 3
4 True 4
5 True 5
6 True 6
Partition 2
7 False 7
8 False 7
9 False 7
10 True 8
11 True 9
In original data timestamps are already sorted (though minor misordering may occur)
My questions are:
Does this sound like a correct approach? If so, how do I perform such a partition? Going through docs, I found set_index, which in my case doesn't help, and a map_overlap function which also doesn't fit since it uses fixed threshold for each partition, which I don't know beforehand.
What would happen if I have a long range of rows with delta==True? I will need to somehow partition it as well, while keeping reference to previous partition until I reach a partition which contains full state, i.e. records with delta==False. I guess this should be handled by a task graph, but is there a way tell Dask to wait for a certain partition to be processed?
Thanks!