azure / rtable Goto Github PK

With the latest changes in repair and write protocols, batch operations were not brought up to speed.

Currently, we dont and fail with conflict

One suggestion is Microsoft.Azure.Labs.RTable

• Detect storage failure. Log critical event when outage is detected.
• If auto-reconfiguration is enabled, any client (preferably elected leader by taking leases on configuration blobs) can reconfigure the chain.
• Similarly, any client should be able to re-introduce a replica once it detects that storage is back online.
• build some kind of hysteresis control.

Currently, clients will need to re-read the data and retry. it is possible for clients to overwrite the data of an in-progress write.

When a view changes, the rtable client must continiue to use the older view till its lease expires. To do this, it must operate on a cached view and check for its validity before each transaction

There are important files that Microsoft projects should all have that are not present in this repository. A pull request has been opened to add the missing file(s). When the pr is merged this issue will be closed automatically.

Microsoft teams can learn more about this effort and share feedback within the open source guidance available internally.

Merge this pull request

The user of rtable library should not be exposed to stale view. For conflict, it should be the same error contract as xTable has for conficts

This isn't necessarily a "bug", but it does need explicitly called out in the documentation.

In ReplicatedTable.cs, there is a fair bit of logic that is deciding whether or not a lock has expired based on the clients clock (via calls to DateTime.UtcNow. The trouble with this is that any client clock which has drifted may inadvertently begin experiencing unexpected behaviors.

For example, in ReplicatedTable.cs lines 715 and 1400, we have the following conditional checks:

if (DateTime.UtcNow >= currentRow._rtable_LockAcquisition + this._configurationWrapper.GetLockTimeout()) {...}

Since rtable runs on the clients, the assumption in this code is that all clients have stable clocks - which, as we know is not a trivial assumption to make.

Continuing the previous example, if we consider a client clock that has drifted ahead. If this case, the client will begin evaluating this condition as True, thereby allowing Flush2PC to proceed when it should have not. Likewise, a client who has drifted behind, will likely never be able to flush writes, since the above condition will end up returning False most of the time.

Apart from documenting the assumption of consistent global time (which needs to be called out), has any thought been given to moving away from wall-time? For example, storing a vector/logical clock in the entity metadata, and implemented standard Lamport-style logic in the clients to determine whether they are behind or not, etc.