Error recovery for LSP creation

[jamra]

Hello,

Hope your holidays are going well.

I was wondering how well SqlGlot can support error recovery. The idea is to lean on this parser for scoping while the SQL author types out their SQL. That way, we can autocomplete with contextual information even though the SQL would currently be in an error state.

If error recovery is not on the radar, I made a parser for Trino with tree-sitter and am wondering if I could leverage SqlGlot as an engine of sorts where I parse my Trino SQL then run the scoping code from SqlGlot. I hope that makes sense.

I will try to open source the parser I have once I get permission to do so, but was wondering if there was a way to collaborate somehow. It could create a baseline for an LSP that others can benefit from.

[tobymao]

you can do it somewhat. you can set the level to IGNORE, which will make the parser much more lenient.

happy to collaborate if it makes sense, just let me know what you need.

[jamra]

I'll give it a try here.

Basically, I wrote a Trino parser using tree-sitter. I'll hopefully open source it once I get permission. Tree-sitter generates a really fast C parser with various bindings in other languages such as python. One of the benefits is that the parser is error recovering, opening up features such as continuing to parse the tree if you have a dangling comma. An example is with the SQL select a, from table_name;. The from in this case will still be parsed to produce tree nodes which could be walked for scope information.

This opens up the use case of writing a proper LSP. I was thinking of writing a few more parsers including one for Presto and Minerva. The SqlGlot code does a great job of separating semantic analysis from tree structure. Perhaps there is a way to wrap the tree-sitter tree in a way where we can reuse the same scoping code. One tertiary benefit is the raw speed of the parser, but the primary goal would be to provide a great developer experience.