Thanks James. Filed https://issues.apache.org/jira/browse/PHOENIX-2240.

On Tue, Sep 8, 2015 at 12:38 PM, James Heather <james.heather@mendeley.com> wrote:
Thanks.

I've discovered that the cause is even simpler. With 100M rows, you get collisions in the primary key in the CSV file. An experiment (capturing the CSV file, and counting the rows with a unique primary key) reveals that the number of unique primary keys is about 500 short of the full 100M. So the upserting is working as it should!

I don't know if there's a way round this, because it does produce rather suspicious-looking results. It might be worth having the program emit a warning to this effect if the parameter size is large, or finding a way to increase the entropy in the primary keys that are generated, to ensure that there won't be collisions.

It's a bit surprising no one has run into this before! Hopefully this script has been run on that many rows before... it seems a reasonable number for testing performance of a scalable database... (in fact I was planning to increase the row count somewhat).

James


On 08/09/15 20:16, James Taylor wrote:
Hi James,
Looks like currently you'll get a error log message generated if a row is attempted to be imported but cannot be (usually due to the data not being compatible with the schema). For psql.py, this would be the client side log and messages would look like this:
            LOG.error("Error upserting record {}: {}", csvRecord, errorMessage);

FWIW, we have a "strict" option for CSV loading (using the -s or --strict option) which is meant to cause the load to abort if bad data is found, but it doesn't look like this is currently checked (when bad data is encountered). I've filed PHOENIX-2239 for this.

Thanks,
James

On Tue, Sep 8, 2015 at 11:26 AM, James Heather <james.heather@mendeley.com> wrote:
I've had another go running the performance.py script to upsert 100,000,000 rows into a Phoenix table, and again I've ended up with around 500 rows missing.

Can anyone explain this, or reproduce it?

It is rather concerning: I'm reluctant to use Phoenix if I'm not sure whether rows will be silently dropped.

James