Collections of Documents on the other hand, offer a much more relaxed approach. Collections are just namespaces where we insert documents. Documents are objects of any schema and format; but in practice, it almost always is JSON. There are no enforced types, no constraints, no guarded references between documents in different collections. In the same collection, we might have documents of completely different schema - flexibility and openness to any data and column types rules here. In tables, rows have columns of simple, scalar types (mostly) - numbers, ids, strings, dates, timestamps and so on. In collections, documents have fields comprising both simple and composite types like arrays and other documents, nested inside. Same field in different documents, but still of the same collection, might have different types as well - almost anything is allowed here.
We are holding off from fully proposing this at this time because
。WPS官方版本下载对此有专业解读
proposal, but we feel that the simplifications promised by the idea
NFAs are cheaper to construct, but have a O(n*m) matching time, where n is the size of the input and m is the size of the state graph. NFAs are often seen as the reasonable middle ground, but i disagree and will argue that NFAs are worse than the other two. they are theoretically “linear”, but in practice they do not perform as well as DFAs (in the average case they are also much slower than backtracking). they spend the complexity in the wrong place - why would i want matching to be slow?! that’s where most of the time is spent. the problem is that m can be arbitrarily large, and putting a large constant of let’s say 1000 on top of n will make matching 1000x slower. just not acceptable for real workloads, the benchmarks speak for themselves here.