Immediately following tablet move operations, version 1.1.1 can exhibit transient snapshot isolation violations in the Jepsen bank test. @manishrjain says that there are known issues with tablet moves in 1.1.1, so this might be fixed already in master. I'm just writing this up so we have a record!
What version of Dgraph are you using?
1.1.1
Have you tried reproducing the issue with the latest release?
1.1.1 is the latest official release; I haven't seen this on 1.1.1-48-g157896305 yet.
What is the hardware spec (RAM, OS)?
A 5-node EC2 m4.large cluster.
Steps to reproduce the issue (command/config used to run Dgraph).
With Jepsen 3bff032adf3a4277e5cbbc2cd05ecec90c69f61e, run
lein run test --version 1.1.1 --concurrency 2n --nemesis move-tablet --time-limit 300 -w bank
Expected behaviour and actual result.
With an initial starting balance of $100 across all accounts, we expect every read to observe $100. However, immediately following move-tablet operations (grey vertical lines), clients can temporarily observe values as low as 45. Because these issues are transient, I think they're probably constrained to the read path; there's no evidence thus far that updates can permanently alter the total amount of money in all accounts.
Immediately following tablet move operations, version 1.1.1 can exhibit transient snapshot isolation violations in the Jepsen bank test. @manishrjain says that there are known issues with tablet moves in 1.1.1, so this might be fixed already in master. I'm just writing this up so we have a record!
What version of Dgraph are you using?
1.1.1
Have you tried reproducing the issue with the latest release?
1.1.1 is the latest official release; I haven't seen this on 1.1.1-48-g157896305 yet.
What is the hardware spec (RAM, OS)?
A 5-node EC2 m4.large cluster.
Steps to reproduce the issue (command/config used to run Dgraph).
With Jepsen 3bff032adf3a4277e5cbbc2cd05ecec90c69f61e, run
lein run test --version 1.1.1 --concurrency 2n --nemesis move-tablet --time-limit 300 -w bank
Expected behaviour and actual result.
With an initial starting balance of $100 across all accounts, we expect every read to observe $100. However, immediately following move-tablet operations (grey vertical lines), clients can temporarily observe values as low as 45. Because these issues are transient, I think they're probably constrained to the read path; there's no evidence thus far that updates can permanently alter the total amount of money in all accounts.
20200109T200137.000Z.zip
The text was updated successfully, but these errors were encountered: