Cassandra: Load large data fast

Question

We're currently working with Cassandra on a single node cluster to test application development on it. Right now, we have a really huge data set consisting of approximately 70M lines of texts that we would like dump into a Cassandra.

We have tried all of the following:

• Line by line insertion using python Cassandra driver
• Copy command of Cassandra
• Set compression of sstable to none

We have explored the option of the sstable bulk loader, but we don't have an appropriate .db format for this. Our text file to be loaded has 70M lines that look like:

2f8e4787-eb9c-49e0-9a2d-23fa40c177a4    the magnet programs succeeded in attracting applicants and by the mid-1990s only #about a #third of students who #applied were accepted.

The column family that we're intending to insert into has this creation syntax:

CREATE TABLE post (
  postid uuid,
  posttext text,
  PRIMARY KEY (postid)
) WITH
  bloom_filter_fp_chance=0.010000 AND
  caching='KEYS_ONLY' AND
  comment='' AND
  dclocal_read_repair_chance=0.000000 AND
  gc_grace_seconds=864000 AND
  index_interval=128 AND
  read_repair_chance=0.100000 AND
  replicate_on_write='true' AND
  populate_io_cache_on_flush='false' AND
  default_time_to_live=0 AND
  speculative_retry='99.0PERCENTILE' AND
  memtable_flush_period_in_ms=0 AND
  compaction={'class': 'SizeTieredCompactionStrategy'} AND
  compression={};

Problem: The loading of the data into even a simple column family is taking forever -- 5hrs for 30M lines that were inserted. We were wondering if there is any way to expedite this as the performance for 70M lines of the same data being loaded into MySQL takes approximately 6 minutes on our server.

We were wondering if we have missed something? Or if someone could point us in the right direction?

Many thanks in advance!

Answer 1

The sstableloader is the fastest way to import data into Cassandra. You have to write the code to generate the sstables, but if you really care about speed this will give you the most bang for your buck.

This article is a bit old, but the basics still apply to how you generate the SSTables .

If you really don't want to use the sstableloader, you should be able to go faster by doing the inserts in parallel. A single node can handle multiple connections at once, and you can scale out your Cassandra cluster for increased throughput.

Answer 2

I have a two node Cassandra 2.? cluster. Each node is I7 4200 MQ laptop, 1 TB HDD, 16 gig RAM). Have imported almost 5 billion rows using copy command. Each CSV file is a about 63 gig with approx 275 million rows. Takes about 8-10 hours to complete the import/per file.

Approx 6500 rows per sec.

YAML file is set to use 10 gigs of RAM. JIC that helps.