Since Couchbase stores metadata information about each document, and the metadata information is stored in RAM. In my case , when the number of records exceed some threshold, Couchbase will use nearly 90% of the data and reject the write request.
Is there anyway to circumvent this problem besides adding more machines?
Not at the moment.
But having the metadata in memory does improve performance significantly especially for get misses if the key doesn’t exist.
We have tunable memory planned for our next major release tentatively planned for the May timeframe. This will allow you to configure whether a bucket should have metadata in memory or not. But as with any of these features, there could be a performance impact based on your use case.
Some users reduce the impact of metadata by grouping more data into fewer documents, but not sure if this is an option for you.