Documentation
¶
Overview ¶
Package atomicops is a DST driver that differentials FoundationDB's ten atomic mutation types (Add / And / Or / Xor / Max / Min / ByteMax / ByteMin / AppendIfFits / CompareAndClear) end-to-end through the SimFDB backend against an INDEPENDENT Go reference model.
Why it earns its keep on top of the existing hunt coverage:
- The interleave driver exercises atomic-ADD only; the other nine ops — and every width / endianness / missing-key edge — are otherwise unprobed by any hunt.
- SimFDB's applyAtomic is already differential-tested against the pure-Go client's applyAtomic (both are ports of C++ Atomic.h). This driver adds a genuinely THIRD, independent reference: the numeric ops (Add/Max/Min) run through a math/big little-endian path instead of a byte carry-loop, so a bug common to BOTH byte-loop ports still surfaces here.
- It drives the ops END-TO-END through the backend — the commit path, the clear→delete wiring for CompareAndClear, and RYW folding inside an open transaction — none of which the pure-function differential touches.
Two absolute oracles (the independent model is the reference; zero violations == agreement):
STATE after each committed transaction a fresh read of the whole keyspace equals the model
— this is where key PRESENCE is checked (absent vs present-empty).
RYW a read of a key inside the open transaction, after its pending atomic ops, equals the
model's in-transaction projection (SimFDB folds RYW via resolveKey).
The load-bearing distinction the model tracks and the STATE oracle probes is ABSENT (no key) vs PRESENT-EMPTY (key present with a zero-length value): AndV2 / MinV2 / ByteMin / CompareAndClear all branch on it. A store that dropped a present-empty value on commit or on read would flip an atomic op onto its absent branch — exactly the divergence this driver is built to catch.
This is RFC-199 Tier 2. It plugs into the shared hunt machinery via hunt.Workload.
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
Types ¶
type Result ¶
type Result struct {
Report *hunt.Report
Txns int
OpsRun int
Counts map[opcode]int // op -> times executed
}
Result is the rich outcome: the standard hunt.Report plus per-op counts so a test can assert every one of the ten atomic ops actually fired (the NO-FAKE-CHECKBOX proof that the driver exercises the whole surface, not just add).
type Workload ¶
type Workload struct {
Keys int // distinct keys (default 6)
Txns int // transactions per run (default 12)
OpsPerTxn int // ops per transaction (default 8)
RYWChecks bool // also assert in-transaction RYW reads (default true)
// EdgeBias skews param widths to {0,1} and Set to empty values, hammering the absent-vs-
// present-empty branches (AndV2 / MinV2 / ByteMin / CompareAndClear) far harder than the
// balanced profile does.
EdgeBias bool
}
Workload is a seeded atomic-op differential over SimFDB. It ignores the record-oriented hunt.Config; its knobs live on the struct.
Source Files