I'm not sure I would use bothE(). Have you tried just outE()?

I would be curious to see if it's querying that's slow or if it's dropping.  Try using profile() or using time of count() instead of drop() if profile() is not available.

Also, what's the frequency of purging?

Assuming the problem is drop speed, I would experiment with finding optimal batch sizes for both g.V().outE() and g.E(). No timestamp filter (unless query speed is negligible), want to isolate batch size so be sure to use a dataset you can blow away.  Obviously 10k is "slow". I would be curious to know how "slow" is slow here. But bisect your way to optimal size.

Then, find the distribution (histogram) over time that closely resembles that batch size. i.e. if you find that the optimal batch size is 5k, then look for how many minutes/hours it takes to make 5k items max.

Let's say it's 1 minute.

Then starting at your oldest time, drop for that 1 minute period on the timestamp filter (begin < timestamp <= end). And simply keep moving that 1 minute period up, drop, repeat until you've reached the timeframe of which should be retained.

See how that goes.  If that proves well, from there you could experiment with parallelism of those chunks. Other possibilities involve hardware (scaling), memory settings, cassandra, ES, etc.

Robert Dale

On Thu, Mar 16, 2017 at 10:37 AM, <k...@...> wrote:

Hi,

I have been using Titan for a while and have recently moved to Janusgraph. Thanks for revitalising the project!

I am using Janusgraph to look at social chat that is generating approximately 7 million messages a day. I want to keep 1 week's worth and when loading today's data, purge the messages that are more than 7 days old.  Are there recommendations for performing this kind of bulk delete of edges?

I am using Cassandra as the backend store and I have messages as edges, with a timestamp property. I have an vertex-centric edgeIndex on the timestamp property and I also have an elasticsearch index on the same timestamp property. I have been running drops in batches of 10,000 and committing the transaction - trying to use (one of) the indexes to do this at speed.

The 2 ways I have tried are:

g.V().bothE('message').has('timestamp', lt(start_date)).limit(10000).drop().iterate()  // uses the edgeIndex

g.E().has('timestamp', lt(start_date)).limit(10000).drop().iterate()                               // uses the elasticsearch index

Both of these are  proving to be very slow, so I am probably missing something?

Thanks for any advice!

Kevin

--
You received this message because you are subscribed to the Google Groups "JanusGraph users list" group.
To unsubscribe from this group and stop receiving emails from it, send an email to janusgraph-use...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.