Semester7

Notes of courses done/attended in semester 7 in college

consistency means txn executing on consistent db leaves it in consistent state, i.e. all constraints are satisfied
in distributed diff meaning
many models
eventually consistent dekha leaderless me
- maybe in inconsistent at some point
- but eventually will come to consistent

about a single opern across various nodes
when system is lineaz, any write is seen as single data write even though theremight be multiple copies
one copy of data abstraction it provides
any changes are visible to subsequent operations
treats all replicas as single data
it is a recency gyuarantee

Compare and Set operations
- when u read a value from replica, assuming u r making change, someone else might have also
- so while making change, u should also give a condition, ki agar value yeh ho, tb yeh kardo
- else, overwrite and lineariz nahi hoga

it is concerned about all operns (agar koi photo change karra woh bhi, faltu me)

only causally related operns must occur in order, others ka idc
eg, blog me reply comes after comment only. kaunsa comment pehle aya does not matter. so reply is related to comment

diff from sequent consis bcz operns from same node with independent causal order can occur in any order
order is only applicable to independent changes

lineariz me total order hai, any opern performed in one node is also in other nodes
there is no concurrent write in lineaz
one copy of data hai, no 2 ppl can write to same data, so no concurrent writes
in causal, order is partial
independent change pe hi order applicable hai, among causal chains, no order
lineariz => causal but not vice versa
lineariz me availability might be harmed, bcz user ko inform karne se pehle we change all replicas
causal does not affect availability provided u can make data in such a way that one causal change can be made in one node in same order, in other nodes it should when they come in contact

can be done using sequence number
har opern ko ek seq numb de de, and value se pata chal jayega ki kaun pehle aya
the seq nums should be globally allocated since kisi bhi node pe ho sakta write opern
single leader replication (all writes go to one leader only) me it makes an order of all operns, same order is in which other applies to their dataet

non causal seq num generators
- odd and even
- ecery node can generate
- pehle globally generate tha
- does not guarantee total order

even-odd, etc do not guarantee causal consistency
bcz seq num are specific to indiv nodes
operns cannot be compared across nodes so , no causality
lamport timestanmps = counter specific to ndoe but node id is attached to counter
- provides total order
- since node counters can be compared
- causally consistent, how?
  - any node which sees seq num higher than itself will make it 1

here counter is being increased
in node 1, it is 1
node 2 me it does writes, count increases
node 1 does a read, gets 5,2 and this counter value (5) is greater than its value (1), so it makes its value as 5
ensures total order and causality
what if client A contacts node 1 after client b writes 3 in node 2
so here count will start increasing, from 1->2->3…
the total order is fine but emerges only after all operns are done.

2 assumptions
- reliable delivery
  - nothing lost
- totally ordered delivery
  - delivered to everybody in same order

in tob, when msgs are originated froma node, order is fixed, others can add to it, but not change it
tob is exactly what we need for db replication
zookeeper me implement hota yeh
tob is same as linearizability

line me every write should appear to others, in tob, we r getting others to agree on an order, consensus = majiority should agree

2 phase commit hota
we have coordinator and participant
coord sends req (can u commit?) to every p, they say yes or no, if they all agree
if all say yes, then commit bhejta coordinator sabko, else rollback

what happens in terms of failure?
say coordinators rcvd yes or no from nodes and when coord sends in second phase abort or commit, and some node is not able to hear it, then node blocks. It cannot abort or commit itself.
so what happens is, after some timeout, it starts sending messages to neighbouring nodes (jinko yeh janta), and if anyone says commit, it commits.
if it is not able to connect with any other node, it blocks then
if coordinator is crashed in second phase, then no body knows and are waiting.
but if rcvd by some, not by some, then prob. bcz if nodes do not know about each other, then bt
so it can end up blocking indefinitely