When a node restarts, it might be canceling one recovery of a shard id only to get another one in the next cycle. We should detect this case and handle it properly.
This is a fix to the annoying message seen by users: suspect illegal state: trying to move shard from primary mode to replica mode.
Two main changes:
Improve cluster resiliency to disconnected sub clusters. If a node pings a master and that node is no longer registered with the master, improve the rejoin process of that node to the cluster. Also, if a master receives a message from another master, pick one to force to rejoin the cluster (based on cluster state versioning).
On quick rolling restart, without waiting for shard allocation, the shard allocation logic can mess up its counts, causing for strange logic in allocating shards, or validation failures on routing table allocation.
Allow to configure store throttling (only applied on file system based storage), which allows to control the maximum bytes per sec written to the file system. It can be configured to only apply while merging, or on all output operations. The setting can eb set on the node level (in which case the throttling is done across all shards allocated on the node), or index level, in which case it only applied to that index.
The node level settings are indices.store.throttle.type to set the type, with values of none, merge and all (defaults to none). And, also, indices.store.throttle.max_bytes_per_sec (defaults to 0), which can be set to something like 1mb.
The index level settings is index.store.throttle.type for the type, with values of node, none, merge, and all. Defaults to node which will use the "shared" throttling on the node level. And, index.store.throttle.max_bytes_per_sec (defaults to 0).
Compressing the stored fields file (the .fdt file) directly allows to have better compression on the size of the index, specifically when indexing (and storing) small documents. The compression will be considerably more effective compared to compressing each doc on its own (when setting compress on the _source mapper). The downside is that more data needs to be uncompressed when loading documents.
The settings to control it is `index.store.compress.stored_fields` set to `true` (it defaults to `false`), and can be enabled dynamically using the update settings API. This allows to enabled compression at a later stage (i.e. old time based indices), and then optimize the index to make sure it gets compressed.
The join proces received on the master from a node joining the cluster now tries to validate by connecting to the relevant node. It should also send a message to it to double check that all is in order.
* added "linc6" as a package dependency as it's required for file
lib/sigar/libsigar-amd64-linux.so and lib/sigar/libsigar-x86-linux.so
Refer to Debian Policy Manual section 8.6 (Dependencies between the
library and other packages - the shlibs system) for details.
Embed the code now in our source, since jsr166e jar generation with 1.6 instead of 1.7 is complicated when doing it on its own as it relies on ThreadLocalRandom, and we have it in jsr166y
