vault backup: 2025-03-24 08:44:22
This commit is contained in:
parent
dc8a303757
commit
63cf9fa686
2 changed files with 16 additions and 2 deletions
|
@ -1,5 +1,6 @@
|
|||
Let us define $op ->_{proc} op'$ to hold whenever there exists a process p that issues both operations with `res[op]` happening before `inv[op']`.
|
||||
|
||||
### Sequential consistency
|
||||
**Def:** a complete history is sequentially consistent if there exists a sequential history $𝑆$ s.t.
|
||||
![[Pasted image 20250324082534.png]]
|
||||
|
||||
|
@ -17,4 +18,17 @@ In isolation, both processes are sequentially consistent.
|
|||
|
||||
However, no total order on the previous 6 operations respects the semantics of a queue:
|
||||
- if p1 receives b' from Q'.deq, we have that Q'.enq(a'), must arrive after Q'.enq(b')
|
||||
- to respect $\to_{proc}$
|
||||
- to respect $\to_{proc}$, also the remaining behavior of p2 must arrive after
|
||||
- hence, Q.enq(a) arrived before Q.enq(b) and so it is not possible for p2 to receive b from its Q.deq.
|
||||
|
||||
Hence, we have two histories that are sequentially consistent but whose composition cannot be sequentially consistent $\to$ **no compositionality!**
|
||||
|
||||
### Serializability
|
||||
(typical notion in databases)
|
||||
|
||||
- instead of processes, we have transactions
|
||||
- consequently, we have also two other kinds of events: `abort(t)` and `commit(t)`
|
||||
- in every history, we have at most one of these events for every transaction
|
||||
- if the history is complete, we must have exactly one of these events for transaction
|
||||
- a sequential history is formed by committed transactions only
|
||||
|
||||
|
|
Loading…
Add table
Add a link
Reference in a new issue