The lock will happen every time an "on duplicate key" clause is executed.
Conceptually it's like a checkAndPut, but it's batched. If two threads
attempt to write the same non-existent row, one of them will get there
first and get the lock, while the other would wait behind it (and
subsequently trigger the "on duplicate key" logic).
On Tue, Aug 27, 2019 at 9:05 AM Jon Strayer wrote:
> Just to make sure I understand, anytime an atomic upsert is executed
> Phoenix translates that into a checkAndPut. Is that correct?
>
> From the Atomic Upsert documentation:
> "On the server-side, when the commit is processed, the row being updated
> will be locked while the current column values are read and the ON
> DUPLICATE KEY clause is executed."
>
> What I'm not clear on is when the lock happens. Does it happen every time
> a upsert with an "on duplicate key" clause is executed or only after the
> duplicate key is detected?
>
> Also, if there are two threads attempting to write to the same
> non-existing row, is there a race condition where neither thread executes
> it "on duplicate key" clause?
>
> On 8/26/19, 7:28 PM, "Josh Elser" wrote:
>
> Out of the box, Phoenix will provide the same semantics that HBase
> does
> for concurrent updates to a (data) table.
>
>
> https://urldefense.com/v3/__https://hbase.apache.org/acid-semantics.html__;!YggOdHcy3_sa5w!7XIAKeE5bdWVHeRnJhaMbEwH9vjnhUzpxgoFvOlQBqYoCWp72RMXHdT636yxH4bg3to$
>
> If you're also asking about how index tables remain in sync, the
> answer
> is a bit more complicated (and has changed in recent versions).
>
> On 8/26/19 2:51 PM, Jon Strayer wrote:
> > How does atomic update work with multiple clients? Assuming that
> there
> > is no matching record to begin with the access won’t be locked. It
> > seems like two threads could write conflicting data since they both
> see
> > no existing record (NER). Is that correct? Or is there something
> that
> > will serialize the writes so that only one of them sees the NER
> state?
> >
>
>
>
