Hi Jools,
what happens in Cassandra with your scenario is the following:
1) insert new record
-> the record is added to Cassandra's dataset (with the given timestamp)
2) delete record
-> a tombstone is added to the data set (with the timestamp of the deletion,
which should be larger than the timestamp in 1), otherwise, the delete
will be lost.
3) insert new record with same key as deleted record
-> the record is added as in 2), but the timestamp should be larger than
the timestamps from both 1) and 2)
When you compact between 2) and 3), the record inserted at 1) will be thrown
away, but the tombstone from 2) will not be thrown away *unless* the tombstone
was created more than GCGraceSeconds (a configuration option) before the
compaction.
If you do not compact, all records and tombstone will be present in Cassandra's
dataset, and each read operation checks which of the records has the highest
timestamp before returning the most current record (or report an error, if the
tombstone
has the highest timestamp).
So whether you compact or not does not make a difference for your scenario,
as long as all replicas see the tombstone before GCGraceSeconds have elapsed.
If that is the case, it is possible that deleted records come alive again,
because
tombstones are deleted before all replicas had a chance to remove the deleted
record.
Your question about concurrently inserting the same key from different clients
is another beast. The simple answer is: don't do it.
The longer answer: either you use some external synchronisation mechanism
(e.g., Zookeeper), or you make sure that all clients use disjoint keys (UUIDs,
or
keys derived from the clients IP address+timestamp, that sort of thing).
For keys representing user accounts or something similar, I would recommend
using an external synchronisation mechanism, because for actions like account
registration latency caused by such a mechanism is usually not a problem.
For data coming in quickly, where the overhead of synchronisation is not
acceptable,
use the UUID variant and reconcile the data on read.
HTH,
Martin
________________________________
From: Jools [mailto:jool...@gmail.com]
Sent: Wednesday, June 09, 2010 10:39 AM
To: user@cassandra.apache.org
Subject: Inserting new data, where the key points to a tombstone record.
Hi,
I've been developing a system against cassandra over the last few
weeks, and I'd like to ask the community some advice on the best way to deal
with inserting new data where the key is currently a tombstone record.
As with all distributed systems, this is always a tricky thing to deal
with, so I though I'd throw it to a wider audience.
1) insert new record.
2) deleted record.
3) insert record with same key as deleted record.
Now I know I can make this work if I flush and compact between 2 and 3.
However, I don't want to rely on a flush and compact and I'd like to code
defensively against this senario, and I've ended up looking up to see if the
key exists, then if it does then I know I can't insert the data. However, if
the key does not exist then I attempt an insert.
Now, here lies the issue. If I have more than one client doing this at
the same time, both trying to insert using the same key. One will succeed and
ones will fail. However neither insert will give me an indication of which one
actually succeeded.
So should an insert against an existing key, or deleted key produce
some kind of exception ?
Cheers,
--Jools
