C ARNEGIE M ELLON U NIVERSITY

D EPARTMENT OF C OMPUTER S CIENCE

15-445/645 – DATABASE S YSTEMS (FALL 2018)
P ROF. A NDY PAVLO

Homework 5 (by Tupac Shakur) – Solutions

Due: Monday Dec 3, 2018 @ 11:59pm

IMPORTANT:
• Upload this PDF with your answers to Gradescope by 11:59pm on Monday Dec 3, 2018.
• Plagiarism: Homework may be discussed with other students, but all homework is to be
completed individually.
• You have to use this PDF for all of your answers.
For your information:
• Graded out of 100 points; 3 questions total
Revision : 2018/12/10 14:04

Question Points Score

Two-Phase Commit 40
Distributed Joins 25
Replication 35
Total: 100

1
15-445/645 (Fall 2018) Homework 5 Page 2 of 8

Question 1: Two-Phase Commit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . [40 points]

Consider a distributed transaction T operating under the two-phase commit protocol. Let N0
be the coordinator node, and N1 , N2 , N3 be the participant nodes.
The following messages have been sent:

time message
1 N0 to N1 : “Phase1:PREPARE”
2 N0 to N2 : “Phase1:PREPARE”
3 N0 to N3 : “Phase1:PREPARE”
4 N2 to N0 : “OK”
5 N1 to N0 : “OK”

Figure 1: Two-Phase Commit messages for transaction T

(a) [10 points] Who should send a message next at time 6 in Figure 1? Select all the possi-
ble answers.
2 N0
2 N1
2 N2
 N3
2 It is not possible to determine
Solution: N3 has to send a response to N0

(b) [10 points] To whom? Again, select all the possible answers.
 N0
2 N1
2 N2
2 N3
2 It is not possible to determine
Solution: N3 has to send a response to N0

(c) [10 points] Suppose that N0 never received the “OK” response from N1 at time 5 in
Figure 1 (the message got dropped due to a hardware failure). Instead, N0 “times out”
after waiting a certain amount of time. What should happen under the two-phase commit
protocol in this scenario?
2 N0 resends “Phase1:PREPARE” to N1
2 N0 resends “Phase1:PREPARE” to all of the participant nodes
2 N0 sends “ABORT” to N1
 N0 sends “ABORT” all of the participant nodes
2 N0 sends “Phase2:COMMIT” all of the participant nodes
2 N1 resends “OK” to N0
2 It is not possible to determine

Question 1 continues. . .
15-445/645 (Fall 2018) Homework 5 Page 3 of 8

Solution: After a timeout, the coordinator (N0 ) will assume that N1 has failed and it
will mark the transaction as aborted. 2PC requires that all participants respond with
“OK”.

(d) [10 points] Suppose that N0 successfully receives all of the “OK” messages from the par-
ticipants from the first phase (i.e., after time 6 in Figure 1). It then sends the “Phase2:COMMIT”
message to all of the participants at time 7 but N2 crashes before it receives this message.
What is the status of the transaction T when N2 comes back on-line?
 T ’s status is committed
2 T ’s status is aborted
2 It is not possible to determine
Solution: Once the coordinator (N0 ) gets a “OK” message from all participants, then
the transaction is considered to be committed even though a node may crash during the
second phase. In this example, N2 would have restore T when it comes back on-line.

Homework 5 continues. . .
15-445/645 (Fall 2018) Homework 5 Page 4 of 8

Question 2: Distributed Joins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . [25 points]

Answer the following questions about performing joins in a distributed database. You can
assume that the DBMS uses a shared-nothing architecture.

A B C C D E
a1 b2 c3 c1 d4 e1
a4 b5 c6 c2 d3 e2
a1 b2 c4 c3 d4 e5
a5 b3 c2 c1 d2 e3
a8 b9 c7 c3 d6 e8
(a) R(A,B,C) (b) S(C,D,E)

Table 1: Sample database

(a) Consider the relations R(A,B,C) and S(C,D,E) shown in Table 1, where attribute S.C is
a foreign key of attribute R.C.
i. [10 points] What is the output of R n S?
2 { (a4,b5,c3), (a4,b5,c3), (a3,b2,c3) }
2 { (a5,b3,c2), (c2,d3,e2), (a1,b2,c3), (c3,d4,e5), (c3,d6,e8) }
2 { (c2,b3,a5), (c2,d3,e2), (c3,d4,e5), (c3,d6,e8) }
2 { (a1,b2), (a4,b5) }
 { (a5,b3,c2), (a1,b2,c3) }
2 { (a5,b2,c3), (a1,b2,c3), (a5,b3,c2) }
2 None of the above

ii. [10 points] What is the output of S n R?

 { (c3,d6,e8), (c3,d4,e5), (c2,d3,e2) }
2 { (c2,d3,e2), (a1,b2,c3), (c3,d6,e8), (c3,d4,e5), (a5,b3,c2) }
2 { (c1,d4,e1), (c2,d3,e2), (c1,d2,e3) }
2 { (c2,d3,e2), (c1,d4,e1), (c3,d6,e8), (c1,d2,e3), (c3,d4,e5) }
2 { (d3,e2), (d6,e8) }
2 { (c2,d3,e2), (c3,d6,e8) }
2 None of the above

Question 2 continues. . .
15-445/645 (Fall 2018) Homework 5 Page 5 of 8

(b) [5 points] In general, is the semijoin operation symmetric for every posssible database?
That is, is the following equation always true for any possible relations R1 and R2?

R1 n R2 =? R2 n R1 (1)
2 Yes
 No
2 It is not possible to determine
Solution: To be written...

Homework 5 continues. . .
15-445/645 (Fall 2018) Homework 5 Page 6 of 8

Question 3: Replication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . [35 points]

Consider a DBMS using active-passive, master-replica replication with multi-versioned con-
currency control. All read-write transactions go to the master node (N ODE A), while read-only
transactions are routed to the replica (N ODE B). You can assume that the DBMS has “instant”
fail-over and master elections. That is, there is no time gap between when the master goes
down and when the replica gets promoted as the new master. For example, if N ODE A goes
down at timestamp 1 then N ODE B will be elected the new master at 2 . Note that this is not
a realistic assumption but we’re using it to simplify the problem setup.
The database has a single table foo(id,val) with the following tuples:

id val
1 aaa
2 bbb

Table 2: foo(id,val)

For each questions listed below, assume that the following transactions shown in Figure 2 are
executing in the DBMS: (1) Transaction #1 on N ODE A and (2) Transaction #2 on N ODE B.
You can assume that the timestamps for each operation is the real physical time of when it was
invoked at the DBMS and that the clocks on both nodes are perfectly synchronized (again, this
is not a realistic assumption).

time operation time operation

1 BEGIN; 2 BEGIN READ ONLY;
2 UPDATE foo SET val = ‘xxx’; 3 SELECT val FROM foo WHERE id = 1;
3 UPDATE foo SET val = ‘yyy’ WHERE id = 1; 4 SELECT val FROM foo WHERE id = 2;
4 UPDATE foo SET val = ‘zzz’ WHERE id = 2; 5 SELECT val FROM foo WHERE id = 2;
5 COMMIT; 6 COMMIT;
(a) Transaction #1 – N ODE A (b) Transaction #2 – N ODE B

Figure 2: Transactions executing in the DBMS.

(a) Assume that the DBMS is using asynchronous replication with continuous log streaming
(i.e., the master node sends log records to the replica in the background after the trans-
action executes them). Suppose that N ODE A crashes at timestamp 4 before it executes
the third UPDATE operation.
i. [10 points] If Transaction #2 is running under SNAPSHOT ISOLATION, what is the
return result of the val attribute for its SELECT query at timestamp 3 ? Select all that
are possible.
 aaa
2 xxx
2 yyy
2 None of the above

Question 3 continues. . .
15-445/645 (Fall 2018) Homework 5 Page 7 of 8

Solution: SNAPSHOT ISOLATION means that the transaction will only see the
versions that were committed before it started. That means at 2 , Transaction #1
has not committed yet so therefore Transaction #2 cannot see any of its versions.
ii. [10 points] If Transaction #2 is running under the READ UNCOMMITTED isolation
level, what is the return result of the val attribute for its SELECT query at times-
tamp 4 ? Select all that are possible.
 bbb
 xxx
2 zzz
2 None of the above
Solution: READ UNCOMMITTED means that it will read any version of the tuple
that exists in the database. But what version of tuple 1 that the transaction will read
depends on whether the master node shipped the log record over before the query is
executed. Since we are doing continuous log shipping, we have no idea. So it could
read the version of the tuple that existed before Transaction #1 started (i.e., “bbb”)
or after Transaction #1 executed the UPDATE query at 2 (i.e., “xxx”). It cannot be
“zzz” because that query never got executed before N ODE A crashed.

(b) [15 points] Assume that the DBMS is using semi-synchronous replication with contin-
uous log streaming. Suppose that both N ODE A and N ODE B crash at exactly the same
time at timestamp 6 after executing Transaction #1’s COMMIT operation. You can assume
that the application was notified that the Transaction #1 was committed successfully.
After the crash, you find that N ODE A had a major hardware failure and cannot boot.
N ODE B is able to recover and is elected the new master.
What are the values of the tuples in the database when the system comes back online?
Select all that are possible.
 { (1,aaa), (2,bbb) }
2 { (1,xxx), (2,bbb) }
2 { (1,xxx), (2,xxx) }
2 { (1,yyy), (2,bbb) }
2 { (1,yyy), (2,xxx) }
 { (1,yyy), (2,zzz) }
2 None of the above
Solution: Semi-synchronous means that the replica only received the log records from
the master but it did not write them to disk. The master sent the notification to the client
that the txn committed but it is only guaranteed to be durable on disk on the master and
not the replica. When the system come back on-line, we don’t know whether the txn
was also flushed to disk on the replica.
Thus, the only two correct states of the database are if Transaction #1 never executed or
if it did execute. The fact that we are doing continuous log shipping doesn’t matter here
because the transaction’s changes are either committed or aborted. There cannot be any

Question 3 continues. . .
15-445/645 (Fall 2018) Homework 5 Page 8 of 8

partial updates to the database.

End of Homework 5