Appy updated HBASE-18432:
> Prevent clock from getting stuck after update()
> Key: HBASE-18432
> URL: https://issues.apache.org/jira/browse/HBASE-18432
> Project: HBase
> Issue Type: Sub-task
> Reporter: Appy
> Assignee: Appy
> Attachments: HBASE-18432.HBASE-14070.HLC.001.patch,
> There were a [bunch of
> (also copied below) with clock getting stuck after call to update() until
> it's own system time caught up.
> PT = physical time, LT = logical time, ST = system time, X = don't care terms
> Core issue:
> - Note that in current implementation, we are passing master clock to RS in
> open/close region request and RS clock to master in the responses. And they
> both update their own time on receiving these request/response.
> - On receiving a clock ahead of its own, they update their own clock to its
> PT+LT, and keep increasing LT till their own ST catches that PT.
> Proposed solution:
> Keep track of skew in clock. And instead of keeping track of physical time,
> always compute it by adding system time and skew.
> On update(), recalculate skew and validate if it's greater than max_skew.
> On toTimestamp(), calculate PT = ST+skew.
> Issues with current approach:
> Problem 1: Logical time window too small.
> RS clock (10, X)
> Master clock (20, X)
> Master --request-> RS
> RS clock (20, X)
> While RS's physical java clock (which is backing up physical component of hlc
> clock) will still take 10 sec to catch up, we'll keep incrementing logical
> component. That means, in worst case, our logical clock window should be big
> enough to support all the events that can happen in max skew time.
> The problem is, that doesn't seem to be the case. Our logical window is 1M
> events (20bits) and max skew time is 30 sec, that results in 33k max write
> qps, which is quite low. We can easily see 150k update qps per beefy server
> with 1k values.
> Even 22 bits won't be enough. We'll need minimum of 23 bits and 20 sec max
> skew time to support ~420k max events per second in worst case clock skew.
> Problem 2: Cascading logical time increment.
> When more RS are involved say - 3 RS and 1 master. Let's say max skew is 30
> HLC Clocks (physical time, logical time): X = don't care
> RS1: (50, 100k)
> Master: (40, X)
> RS2: (30, X)
> RS3: (20, X)
> [RS3's ST behind RS1's by 30 sec.]
> RS1 replies to master, sends it's clock (50,X).
> Master's clock (50, X). It'll be another 10 sec before it's own physical
> clock reaches 50, so HLC's PT will remain 50 for next 10 sec.
> Master --> RS2
> RS2's clock = (50, X).
> RS2 keeps incrementing LT on writes (since it's own PT is behind) for few
> seconds before it replies back to master with (50, X+ few 100k).
> Master's clock = (50, X+ few 100k) [Since master's physical clock hasn't
> caught up yet, note that it was 10 seconds behind, PT remains 50.].
> Master --> RS3
> RS3's clock (50, X+few 100k)
> But RS3's ST is behind RS1's ST by 30 sec, which means it'll keep
> incrementing LT for next 30 sec (unless it gets a newer clock from master).
> But the problem is, RS3 has much smaller LT window than actual 1M!!
> Problem 3: Single bad RS clock crashing the cluster:
> If a single RS's clock is bad and a bit faster, it'll catch time and keep
> pulling master's PT with it. If 'real time' is say 20, max skew time is 10,
> and bad RS is at time 29.9, it'll pull master to 29.9 (via next response),
> and then any RS less than 19.9, i.e. just 0.1 sec away from real time will
> die due to higher than max skew.
> This can bring whole clusters down!
> Problem 4: Time jumps (not a bug, but more of a nuisance)
> Say a RS is behind master by 20 sec. On each communication from master, RS
> will update its own PT to master's PT, and it'll remain that till RS's ST
> catches up. If there are frequent communication from master, ST might never
> catch up and RS's PT will actually look like discrete time jumps rather than
> continuous time.
> For eg. If master communicated with RS at times 30, 40, 50 (RSs corresponding
> times are 10, 20, 30), than all events on RS between time [10, 50] will be
> timestamped with either 30, 40 or 50.
This message was sent by Atlassian JIRA