Alex Chi Z.

> * (2) Doesn't support fast-path concurrent read. When there are multiple queries accessing the same block, and the block is not present in memory, the `Column` struct will fetch...

We can follow the Postgres convention. IIRC, it will use i64 to sum i32 and use Decimal to sum i64. I think we can start working on it when we...

From my perspective, it's better to do either of the following: 1. `update(&self, array: &ArrayImpl, visibility: impl Iterator)` 2. `update_single(&self, array: &ArrayImpl, idx: usize)` @pleiadesian What do you think?

By the way, for every aggregate state, we need to rewrite the same logic twice: once for batch update, and once for single update... Any idea on simplifying this, or...

> But the cost of the unified interface is that we have to construct a bitmap every time we want to update the aggregate state. Actually not. As you can...

After a second thought, evan a visibility `Iterator` might require `O(N^2)` time complexity in the worst case, when all group keys are different. We need to call `update` N times,...

> how to achieve this? make a snapshot of each epoch and only track the different from last? Exactly, but I didn't come up with a detailed way to achieve...

No worries 🤪

Thanks! By the way, currently there are no way to test if an executor's schema is consistent with its data chunk outputs. So you may simply change the schema (according...

Thanks! The core logic lies in https://github.com/risinglightdb/risinglight/blob/141a2624a527e346900a8c2d185d8176c89e460d/src/storage/secondary/block/block_index_builder.rs#L62-L75 I think we may need to change the encoding of block headers (e.g., put block type and checksum at the end?) in order...