feat: partition table query optimize

[zealchen]

Rationale

Close #1441

Detailed Changes

TLDR

The performance issue with inlist queries is due to the extra overhead from bloom-filter-like directory lookups when scanning each SST file for rows. The solution is to create a separate predicate for each partition, containing only the keys relevant to that partition. Since the current partition filter only supports BinaryExpr(Column, operator, Literal) and non-negated InList expressions, this solution will address only those specific cases.

Changes

1. During the scan building process, when identifying the partitions for a query, we create a PartitionedFilterKeyIndex variable to store the predicate key indices for each expression.
2. In the compute_partition_for_keys_group function, we use a HashMap<partition_id, HashMap<filter_index, BTreeSet<key_index>>> to record the indices of keys involved in partition computation for each group.
3. In the partitioned_predicates function, we construct the final predicates for each partition.
4. In resolve_partitioned_scan_internal, we generate separate requests for each partition.

e.g.
conditions:

1. table schema: col_ts, col1, col2, in which col1 and col2 are both keys,
   and with two partitions
2. sql: select * from table where col1 = '33' and col2 in ("aa", "bb",
   "cc", "dd")

partition expectations:
yield two predicates
p0: col1 = '33' and col2 in ("aa", "bb", "cc");
p1: col1 = '33' and col2 in ("dd")

Other issues discovered

When the inlist key args length is less than three, Expr will be refactored to nested BinaryExpr which bypasses the FilterExtractor.

e.g.
SQL: select * from table where col1 in ("aa", "bb") and col2 in (1,2,3,4,5...1000)
Since ("aa", "bb") has fewer than three elements, the col1 key filter is not included in partition computation, which interrupts the partitioning process in the get_candidate_partition_keys_groups function, as contains_empty_filter is set to true.

Test Plan

1. UT: test_partitioned_predicate
2. Manual test.

[jiacai2050]

Our codebase should prefer absolute import over super import.

[jiacai2050]

horaedb/src/common_types/src/datum.rs

Line 1313 in e065ba7

impl<'a> std::hash::Hash for DatumView<'a> {

DatumView already impl Hash

[zealchen]

Only the Hash trait is not adequate, since a HashSet requires that the elements implement the Eq and Hash traits. Why don't you implement Eq, btw?

[zealchen]

This BTreeSet in PartitionedFilterKeyIndex is used to deduplicate the inlist key value that contributes to the partition calculation.

[jiacai2050]

LGTM

[jiacai2050]

👍 for comments, it make the code more readable!