Optimize dedup to avoid oom

[coolderli]

• distinct() will storage the data, and the downstream will read from the shuffle. We do not need the cache any more.
• Use the count() to instead the collect() to avoid the drive OOM.