Faster workspace execution

[darabos]

I'm addressing multiple issues here:

• Persistent outputs are slower because they write to disk. It's just a bad default.
• The default LIMIT is annoying and no longer necessary if we don't persist the table. (Remove "limit 10" from the default SQL queries #151)
• Each SQL box created 6 VertexToEdgeAttribute operations per input vertex attribute. With 100 attributes and 20 SQL boxes this was 12,000 operations on every setAndGetWorkspace request. That takes around 3 seconds. That's long enough that you start clicking on more things, creating an increasing backlog of requests. In my test I got 20–30 second response times easily.

For the third issue 4 of the 6 VertexToEdgeAttribute operations were unnecessarily caused by the defaultTableName code. That's easy to fix. But the other two are unavoidable if we execute the box.

So I've added a "BoxCache" to avoid unnecessary box executions. This is a huge improvement. Even if you only moved a box, we used to execute all the boxes in the workspace. But with the cache we don't execute anything and it's super fast. When you change something, we only execute that one box.

I went with this universal solution rather than try to improve the SQL box code, because it's not the only box that scales linearly with the number of attributes. E.g. a filter box will pull over all attributes. It's a bit faster than the SQL box, but I still saw 1–2 second latencies with 100 attributes and 20 filter boxes. The cache fixes this for all boxes.

Do you see any potential issues with it?

[xandrew]

This is cool! I like that you don't try to rebuild a new layer of lineage based caching!

[darabos]

Thanks!