PidokuInfra
Glossary

Glossary

Short definitions of every term the course uses. The lesson in brackets is where it is explained.

TermMeaning
AlignmentThe address multiple at which a value must start; the cause of struct padding. (II.04)
ArenaA region of the heap reserved from the OS (III.03); also a pattern of allocating many objects from one block and freeing them together. (III.05)
AtomicAn operation on one word that completes indivisibly. (IV.04)
AttentionThe transformer operation in which a token gathers information from earlier tokens. (VI.05)
BackpressureA slow consumer slowing its producer, usually through a bounded queue. (IV.06)
Bounds checkThe run-time test that an index is within a slice; removable when provably safe. (V.02)
BPEByte-pair encoding: a tokenizer built by repeatedly merging the most frequent adjacent pair. (VI.04)
Cache lineThe 64-byte unit in which the CPU loads and owns memory. (V.03)
cgoGo’s mechanism for calling C. (V.04)
ChannelA typed, synchronized queue between goroutines. (IV.02)
ClosureA function value that captures variables from its surroundings. (I.03)
ContextA value carrying cancellation, a deadline and request-scoped data down a call chain. (IV.03)
Data raceUnsynchronized concurrent access to a variable, at least one a write. (IV.05)
DevirtualizationTurning an interface method call into a direct call. (V.02)
Dynamic batchingGrouping requests that arrive close together into one model call. (VI.08)
EmbeddingA vector representing an item so that similar items are close together. (VI.07)
Escape analysisThe compiler pass deciding whether a value can live on the stack. (III.02)
False sharingIndependent variables on one cache line slowing each other’s writers. (V.03)
G, M, PGoroutine, OS thread, and scheduling context. (IV.01)
GenericsFunctions and types with type parameters constrained by interfaces. (II.06)
GOGCThe heap growth percentage that triggers the next garbage collection. (III.04)
GOMAXPROCSThe number of Ps: how many goroutines run Go code at once. (IV.01)
GOMEMLIMITA soft ceiling on the runtime’s total memory. (III.04)
GoroutineA function running concurrently on its own small, growable stack. (IV.01)
Green TeaThe span-oriented garbage collector implementation, default since Go 1.26. (III.04)
Happens-beforeThe ordering relation that defines when one goroutine’s write is visible to another. (IV.05)
HeapMemory that outlives a function call, managed by the allocator and the GC. (III.01)
HNSWA graph-based approximate nearest-neighbour index. (VI.07)
InliningReplacing a call with the callee’s body. (V.02)
Interface valueTwo words: a type (or method table) and a data pointer. (II.05)
IteratorA function that can be ranged over, yielding values one at a time. (II.06)
IVFAn approximate index that searches only the partitions nearest the query. (VI.07)
KV cacheStored attention keys and values for previous tokens. (VI.05)
LogitsThe model’s raw scores, one per vocabulary entry. (VI.05)
mcache / mcentral / mheapThe allocator’s per-P, per-size-class and global tiers. (III.03)
MCPModel Context Protocol: a standard for exposing tools and data to models. (VI.09)
Method setThe methods available on a type; decides interface satisfaction. (I.04)
ModuleA versioned tree of packages with a go.mod file. (I.05)
NetpollerThe runtime component that parks goroutines on network I/O and wakes them. (V.05)
PGOProfile-guided optimization using a production CPU profile. (V.02)
PreemptionThe runtime interrupting a long-running goroutine. (IV.01)
RecallThe fraction of true nearest neighbours an approximate search returns. (VI.07)
RuneA Unicode code point; int32. (II.02)
Size classOne of the fixed object sizes the allocator serves. (III.03)
Slice headerPointer, length and capacity describing a window on an array. (II.01)
SpanA run of 8 KB pages dedicated to one size class. (III.03)
SSEServer-Sent Events: the streaming format used by LLM APIs. (V.05, VI.06)
StackPer-goroutine memory for function frames; grows by copying. (III.01)
Swiss tableThe hash-table design behind Go’s maps since 1.24. (II.03)
sync.PoolA cache of reusable temporary objects, emptied by the GC. (III.05)
TensorA flat array of numbers plus a shape. (VI.02)
Tool callA model’s request that the program run a named function with JSON arguments. (VI.09)
Tri-colour markingThe white/grey/black algorithm the collector uses to find live objects. (III.04)
Work stealingAn idle P taking runnable goroutines from another P’s queue. (IV.01)
Write barrierCode run on pointer writes during GC marking to keep it correct. (III.04)
Zero valueThe default every variable starts with: 0, "", false, nil. (I.02)

↑↓ navigate↵ openesc close