Berkeley DB Reference Guide: Clock Skew

Berkeley DB Reference Guide:
Berkeley DB Replication

Clock Skew

Since master leases take into account a timeout that is used across all sites in a replication group, leases must also take into account any known skew (or drift) between the clocks on different machines in the group. The guarantees provided by master leases take clock skew into account. Consider a replication group where a client's clock is running faster than the master's clock and the group has a lease timeout of 5 seconds. If clock skew is not taken into account, eventually, the client will believe that 5 seconds have passed faster than the master and that client may then grant its lease to another site. Meanwhile, the master site does not believe 5 seconds have passed because its clock is slower, and it believes it still holds a valid lease grant. For this reason, Berkeley DB compensates for clock skew.

The master of a group using leases must account for skew in case that site has the slowest clock in the group. This computation avoids the problem of a master site believing a lease grant is valid too long. Clients in a group must account for skew in case they have the fastest clock in the group. This computation avoids the problem of a client site expiring its grant too soon and potentially granting a lease to a different site. Berkeley DB uses a conservative computation and accounts for clock skew on both sides, yielding a double compensation.

The DB_ENV->rep_set_clockskew method takes the values for both the fastest and slowest clocks in the entire replication group as parameters. The values passed in must be the same for all sites in the group. If the user knows the maximum clock drift of their sites, then those values can be expressed as a relative percentage. Or, if the user runs an experiment then the actual values can be used.

For example, suppose the user knows that there is a maximum drift rate of 2% among sites in the group. The user should pass in 102 and 100 for the fast and slow clock values respectively. That is an unusually large value, so suppose, for example, the rate is 0.03% among sites in the group. The user should pass in 10003 and 10000 for the fast and slow clock values. Those values can be used to express the level of precision the user needs.

An example of an experiment a user can run to help determine skew would be to write a program that started simultaneously on all sites in the group. Suppose, after 1 day (86400 seconds), one site shows 86400 seconds and the other site shows it ran faster and it indicates 86460 seconds has passed. The user can use 86460 and 86400 for their parameters for the fast and slow clock values.

Since Berkeley DB is using those fast and slow clock values to compute a ratio internally, if the user cannot detect or measure any clock skew, then the same value should be passed in for both parameters, such as 1 and 1.