pgbench regression test failure

francolin just showed a non-reproducing failure in the new pgbench tests:
https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=francolin&dt=2017-09-12%2014%3A00%3A02

The relevant part of the log is

# Running: pgbench -T 2 -P 1 -l --log-prefix=001_pgbench_log_1 --aggregate-interval=1 -S -b se(at)2 --rate=20 --latency-limit=1000 -j 2 -c 3 -r
ok 198 - pgbench progress status (got 0 vs expected 0)
ok 199 - pgbench progress stdout /(?^:type: multiple)/
ok 200 - pgbench progress stdout /(?^:clients: 3)/
ok 201 - pgbench progress stdout /(?^:threads: 2)/
ok 202 - pgbench progress stdout /(?^:duration: 2 s)/
ok 203 - pgbench progress stdout /(?^:script 1: .* select only)/
ok 204 - pgbench progress stdout /(?^:script 2: .* select only)/
ok 205 - pgbench progress stdout /(?^:statement latencies in milliseconds)/
ok 206 - pgbench progress stdout /(?^:FROM pgbench_accounts)/
ok 207 - pgbench progress stderr /(?^:vacuum)/
ok 208 - pgbench progress stderr /(?^:progress: 1\b)/
ok 209 - number of log files
ok 210 - file name format
not ok 211 - transaction count for 001_pgbench_log_1.31583 (3)

# Failed test 'transaction count for 001_pgbench_log_1.31583 (3)'
# at t/001_pgbench_with_server.pl line 438.
ok 212 - transaction format for 001_pgbench_log_1
ok 213 - transaction count for 001_pgbench_log_1.31583.1 (2)
ok 214 - transaction format for 001_pgbench_log_1
ok 215 - remove log files

Apparently, one of the threads ran 3 transactions where the test script
expects it to run at most 2. Is this a pgbench bug, or is the test
being overoptimistic about how exact the "-T 2" cutoff is?

regards, tom lane