Jamie Brandon

Sup :) ``` clojure (deftype Foo [x]) (clojure.core/let [remaining8730 (Foo. nil)] (proteus/let-mutable [y nil] (.x remaining8730))) => nil (clojure.core/let [remaining8730 (Foo. nil)] (proteus/let-mutable [x nil] (.x remaining8730))) => java.lang.IllegalArgumentException: No...

``` julia> data = @from t in train begin @where t.item_nbr == 1 @select t end 0x6 query result id │ date │ store_nbr │ item_nbr │ unit_sales │ onpromotion...

Improve representation of IR and metadata

1

Current design choices * Use `Let` to express sharing * Store metadata (types, keys) in `RelationExpr::Get` rather than look up the corresponding `Let` * Recalculate metadata for other RelationExpr on...

* [x] binary_search * [ ] bloom_filter * [x] no false negatives * [ ] test false positive rate (requires calculating the expected rate, which is hard without poisson distribution...

Optimize hashing and merging

14

~Under a sustained load of transfers from a single client, TB spends 30% cpu in blake3 and 30% in compaction.cpu_merge*. While some of this is due to doing too many...

Prefetch becomes a bottleneck as database size grows

2

I ran a benchmark inserting 500m transfers and saw throughput slow to 8k t/s by the end. I just reran with 60m transfers and tracy enabled (full trace [here](https://drive.google.com/file/d/1UFUDjeRfXcLSWlpFLcNEtdM5RvfFaFvF/view?usp=sharing)). Here...

Compaction reads seem to entirely miss the grid cache.

8

Here's a sample trace zoomed in on a single compact. We see 17 hits and 9008 misses.  Compare to prefetching, with 18260 hits and 9 misses.  Possibly what...

Under heavy load there is a gap after every checkpoint where the server is just waiting on the network.  @sentientwaffle commented: > During a checkpoint, not only can we...

Sorting table_mutable starves IO

4

Most of the time during compaction we're keeping the io queue pretty full. But when one compaction finishes and we sort the mutable table, this occupies the cpu for a...