optimize C API to reduce memory allocations and using PinnableSlice for zero-copy reads

[zaidoon1]

Problem

The current C API implementation has inefficiencies that impact performance in production environments:

1. Double allocations in Get operations: Values are first copied into a std::string, then copied again into a malloc'd buffer
2. Unnecessary string temporaries: Using std::string as intermediate storage adds allocation/deallocation overhead
3. No zero-copy read path: All read operations require at least one allocation and copy
4. Redundant operations: CopyString performed unnecessary sizeof(char) multiplication

Solution

1. Use PinnableSlice for Get Operations

• Before: DB::Get() → std::string → malloc'd buffer (2 allocations, 2 copies)
• After: DB::Get() → PinnableSlice → malloc'd buffer (1 allocation, 1 copy)
• Impact: 50% reduction in allocations and copies

2. Optimize CopyString Helper

• Removed redundant sizeof(char) multiplication
• Single implementation using Slice parameter (works with all types via implicit conversion)
• Added inline for better optimization

3. New Zero-Copy API Functions

Added high-performance alternatives for allocation-sensitive workloads:

• rocksdb_get_pinned_v2/ rocksdb_get_pinned_cf_v2 - Zero-copy read access
• rocksdb_get_into_buffer/ rocksdb_get_into_buffer_cf - Copy into user-provided buffer
• rocksdb_pinnable_handle_* - Handle management functions

Performance Improvements

Operation	Allocations	Improvement
rocksdb_get	2 → 1	50% reduction
rocksdb_get_cf	2 → 1	50% reduction
rocksdb_multi_get (per key)	2 → 1	50% reduction
rocksdb_transaction_get	2 → 1	50% reduction
rocksdb_writebatch_wi_get_from_batch	2 → 1	50% reduction
rocksdb_get_pinned_v2 (new)	0	100% reduction

Functions Optimized (30+)

• All Get variants (regular, CF, with timestamps)
• All MultiGet variants
• All Transaction Get/MultiGet operations
• All WriteBatch Get operations
• KeyMayExist operations
• Metadata getters (column family names, SST file keys, transaction names, DB identity)

Testing

• Added tests for new zero-copy functions
• Added tests for previously untested functions rocksdb_column_family_handle_get_name, rocksdb_transaction_get_name

Migration Path

Applications can adopt improvements in three ways:

1. No changes needed - Existing code automatically benefits from 50% allocation reduction
2. Incremental adoption - Replace hot-path calls with zero-copy variants
3. Full optimization - Use rocksdb_get_into_buffer

[zaidoon1]

@cbi42 would be great if you can take a look at this when you can. Thanks!

[zaidoon1]

@cbi42 bumping this, can you please take a look. Also cc @pdillinger

[zaidoon1]

@cbi42 did you get a chance to take a look at this? What do you think? If you think it's a bad change or won't work as I expect, please let me know

[zaidoon1]

@pdillinger what do you think about this change? is it worth doing?