Improper configuration is by far the largest problem we see people trying to deploy Apache Druid (incubating). The example configurations listed in the tutorials are designed for a small volume of data where all processes are on a single machine. The configs are extremely poor for actual production use.
[Broker processes](../design/broker.html) uses the JVM heap mainly to merge results from Historicals and real-times. Brokers also use off-heap memory and processing threads for groupBy queries. We recommend 20G-30G of heap here.
[Historical processes](../design/historical.html) use off-heap memory to store intermediate results, and by default, all segments are memory mapped before they can be queried. Typically, the more memory is available on a Historical process, the more segments can be served without the possibility of data being paged on to disk. On Historicals, the JVM heap is used for [GroupBy queries](../querying/groupbyquery.html), some data structures used for intermediate computation, and general processing. One way to calculate how much space there is for segments is: memory_for_segments = total_memory - heap - direct_memory - jvm_overhead. Note that total_memory here refers to the memory available to the cgroup (if running on Linux), which for default cases is going to be all the system memory.
[Coordinator processes](../design/coordinator.html) do not require off-heap memory and the heap is used for loading information about all segments to determine what segments need to be loaded, dropped, moved, or replicated.
Any Druid process that handles queries (Brokers, Peons, and Historicals) uses two kinds of direct memory buffers with configurable size: processing buffers and merge buffers.
Each processing thread is allocated one processing buffer. Additionally, there is a shared pool of merge buffers (only used for GroupBy V2 queries currently).
Other sources of direct memory usage include:
- When a column is loaded for reading, a 64KB direct buffer is allocated for decompression.
- When a set of segments are merged during ingestion, a direct buffer is allocated for every String typed column, for every segment in the set to be merged. The size of these buffers are equal to the cardinality of the String column within its segment, times 4 bytes (the buffers store integers). For example, if two segments are being merged, the first segment having a single String column with cardinality 1000, and the second segment having a String column with cardinality 500, the merge step would allocate (1000 + 500) * 4 = 6000 bytes of direct memory. These buffers are used for merging the value dictionaries of the String column across segments. These "dictionary merging buffers" are independent of the "merge buffers" configured by `druid.processing.numMergeBuffers`.
A useful formula for estimating direct memory usage follows:
The `+1` is a fuzzy parameter meant to account for the decompression and dictionary merging buffers and may need to be adjusted based on the characteristics of the data being ingested/queried.
The intermediate computation buffer specifies a buffer size for the storage of intermediate results. The computation engine in both the Historical and Realtime processes will use a scratch buffer of this size to do all of their intermediate computations off-heap. Larger values allow for more aggregations in a single pass over the data while smaller values can require more passes depending on the query that is being executed. The default size is 1073741824 bytes (1GB).
Server maxSize sets the maximum cumulative segment size (in bytes) that a process can hold. Changing this parameter will affect performance by controlling the memory/disk ratio on a process. Setting this parameter to a value greater than the total memory capacity on a process and may cause disk paging to occur. This paging time introduces a query latency delay.
