Ple Row Store

class PleRowStore

Embedding table with resident or synchronous SSD-to-GPU cached storage. The stream outlives this store, and the store outlives its leases. All engine consumers use this stream and release leases only after their GPU work completes.

Public Types

enum class Residency

Values:

enumerator kCACHED

enumerator kRESIDENT

Public Functions

PleRowStore(

PleTable table,

int32_t slotCount,

cudaStream_t stream,

PleIoConfig io = {},

Residency residency = Residency::kCACHED,

std::shared_ptr<PleIoBudget> ioBudget = nullptr

)

Resident mode requires slotCount == rowCount + 1 and finishes bounded preload before returning.

~PleRowStore() noexcept

PleRowStore(PleRowStore const&) = delete

PleRowStore &operator=(PleRowStore const&) = delete

std::vector<RequestLease> prepare(

std::vector<std::vector<int64_t>> const &requests

)

Tensor &pool()

inline Stats const &stats() const noexcept

Counters for the most recent admitted prepare, including partial reads on failure.

inline Stats const &preloadStats() const noexcept

class RequestLease

Ordered row mapping and optional cache pins. The store outlives this handle.

Public Functions

RequestLease(RequestLease const&) = delete

RequestLease &operator=(RequestLease const&) = delete

RequestLease(RequestLease &&other) noexcept

RequestLease &operator=(RequestLease &&other) noexcept

std::vector<int32_t> const &mapping() const

void release() noexcept

struct Resources

Public Functions

inline Resources(

PleTable const &table,

int32_t slots,

std::unique_ptr<PleIoBudget::Reservation> reservation

)

Public Members

Tensor pool

std::unique_ptr<Staging> staging

struct Staging

Public Functions

inline Staging(

int64_t rowBytes,

std::unique_ptr<PleIoBudget::Reservation> reservation

)

inline ~Staging()

Public Members

std::unique_ptr<PleIoBudget::Reservation> budget

Tensor host

Tensor device

std::vector<PleFileHandle> files

std::vector<int64_t> shardEnds

std::unique_ptr<PleReadExecutor> reader

struct Stats

Public Members

size_t uniqueRows = {}

Computed for cached admission; resident queries do not deduplicate.

size_t misses = {}

size_t readCalls = {}

size_t usefulBytes = {}

Requested miss payload, excluding page amplification.

size_t readBytes = {}

Bytes returned by pread; not physical SSD traffic for buffered I/O.

size_t fillWaves = {}

class RequestLease

Ordered row mapping and optional cache pins. The store outlives this handle.

Public Functions

RequestLease(RequestLease const&) = delete

RequestLease &operator=(RequestLease const&) = delete

RequestLease(RequestLease &&other) noexcept

RequestLease &operator=(RequestLease &&other) noexcept

std::vector<int32_t> const &mapping() const

void release() noexcept

struct Stats

Public Members

size_t uniqueRows = {}

Computed for cached admission; resident queries do not deduplicate.

size_t misses = {}

size_t readCalls = {}

size_t usefulBytes = {}

Requested miss payload, excluding page amplification.

size_t readBytes = {}

Bytes returned by pread; not physical SSD traffic for buffered I/O.

size_t fillWaves = {}