Couchbase Server 3.0

Performance Recap

Agenda

• Different perspectives
• Main achievements
• Unmet expectations

KV cases

• Faster persistence
• Faster replication
• Faster compaction

• Slower read performance
• More tuning challenges
• Memory issues were not resolved

But...

Would I buy/upgrade?

Maybe...

but it looks more like aggressively tuned 2.5

Rebalance and Failover

• KV rebalance is faster
• Delta recovery is even faster
• Graceful failover is... graceful

But...

• It doesn't scale beyond "Giga" items
• Delta recovery is totally unpractical

Would I buy/upgrade?

Yes...

but mostly because of "graceful" failover

Heavy DGM

• Full ejection attempts to address sizing problem

But...

• Impact on other API
• "Storage" doesn't scale anyways
• Errors and timeouts are not uncommon
• Safe level is still around 5%

Would I buy/upgrade?

Yes...

if I have millions/billions of small documents

Views

• Lower end-to-end latency
• Faster indexing
• Better consistency

But...

• Worse "stale=update_after" latency and throughput
• "stale=false" queries still take several seconds
• Rebalance with views is 10-15x slower than KV-only
• Erlang introduces too many problems
• Disk based indexing has troubles when data size increases

Would I buy/upgrade?

No...

unless persistence latency was my the only problem

XDCR

• Lower end-to-end latency in LAN setups
• Higher replication throughput
• More control over replication speed

But...

• Speed vs. resources trade-off
• Still struggling with architecture limitations in WAN

Would I buy/upgrade?

Yes...

at least I will lose my data less frequently

Multitenancy

• Shared thread pool
• No mccouch
• Bucket priorities

But...

• No specific architecture changes

Would I buy/upgrade?

Yes...

if I have couple dozen buckets

The most frustrating

• Delta recovery
• R16
• View queries

The most promising

• Full ejection
• Shared thread pool