Raven’s Storage: Building a Sorted String Table
Applying the lessons from the leveldb codebase isn’t going to be a straightforward task. I already discussed why we can’t just use leveldb directly, so I won’t go into that.
I decided to see if I can implement a leveldb inspired storage for RavenDB. And the first thing that I wanted to try is to build an SST. That seems like an easy first step. SST are just sorted data in a file, and it require nothing much to write to them. As mentioned, we can’t just do a line by line port of the code. As easy as that would be. The way leveldb manages memory is… not really suitable for what we can / should do in .NET.
- I want this to be an efficient .NET implementation.
- It is inspired, but not intended to be compatible with leveldb.
The leveldb API deals with byte arrays all over the place. This makes sense for a C++ codebase, but it is horrible for a .NET application, especially when we expect a lot of our values to be large enough to hit the LOH. That means fragmentation, and pain down the road. Instead, our API uses ArraySegment<byte> for keys, and Stream for values. Since keys are expected to be relatively small, I don’t foresee this being a problem. And the values are streams, so they are easily handled without introducing any cost from the API.
Another thing that leveldb does quite a lot is batch things in memory for a while. It may be the current block, it may be the current data block, it may be the index block, but it does so quite often. That works nicely for C++ apps with expected small values, but not so much for our expected use case. So I want to avoid as much as possible holding items in managed memory. Here is the API for creating an SST:
As you can see, we uses two streams here, one to actually write to the table, the second is a temporary stream that we use to write the index block while we are working, then we merged it back to the actual sst file. Note that after building the table, the temp file can be deleted (indeed, we marked is as delete on close, so that would automatically happen).
That part was easy, all it required was simple I/O for generating the file. The more interesting part is going to be reading the values out.