Implementation is completed on piha. Clean-up and get changes committed.

currently all branches built by travis-ci pull from zlog master

Much of the implementation is currently in the header and we'd like to move that into the libzlog/backend/ceph.cc file.

We can use N threads in libseqr to talk to the sequencer, and then queue up requests from multiple clients of the library instance for batching.

http://stackoverflow.com/questions/15073524/boost-asio-single-threaded-performance

There are a few different executables in src/kvstore used to stress test the kvstore library, but there are no basic unit tests that can be run as smoke tests, especially for adding tests for specific scenarios and semantics.

This issue would add a new google test case for the key-value store library. The goal would be cover the basic operations (get, put, transactions, iterators), and serving as a place for new tests to be added when needed.

This project currently uses autotools and cmake for build systems, and each library and executable is defined in both frameworks. The only major difference between the two is that the zlog java bindings are built with autotools and not cmake.

This issue would (1) fixup the cmake files to build the java bindings (or potentially use some other solution (eg maven) to build the java bindings), and (2) remove the autotools framework so that only cmake (and possibly a java-specific build system) is used.

use cases: zero-copy straight to bufferlist for zlog backend, or serialize to std::string for the vector-based testing backend.

This is just a basic task that ensures that header files and paths are all setup well. It should be as easy as changing the ci run scripts to do:

1. cmake .; make
2. mkdir build; cd build; cmake ..; make

This depends on the C api support backends in #85

See 438ef53

The google test framework is used to run unit tests in this project. The entire google test framework was imported into this project, but this makes it hard to update the software, creating massive commits that complicate the git history.

This issue would (1) remove the directory src/gtest, (2) replace it with a git submodule pointing at a release tag of the google test framework git repository, and (3) wire up the build system and tests.

Currently the C bindings are associated with the ceph backend. This needs to be generalized so we can also pass in a backend to the C bindings. The easiest thing to do is probably give each backend a C API as well that returns an opaque backend object and then pass that to the C api.

This issue will switch the datatype for key-value storage to use a blob Slice type.

The cmake build of zlog contains hooks for building with code coverage support. In travis-ci we can run tests with coverage enabled and upload the coverage results to https://coveralls.io/.

Check for in-flight reconfigurations and block waiting for completion to avoid thundring hurd. note this only help among users of a zlog client and not across client processes.

This is important because as it stands now even the RAM backend needs to link to libseqr which is a pain.

see cmake/modules/CodeCoverage.cmake

It probably shouldn't do this.

I think we can keep the source in a doc directory and use git subtree publish to push into gh-pages. We'll need to find a new location for the java doc (e.g. /api/java/).

Placeholder for a sequencer failover / recovery solution.

The key-value store library contains a backend abstraction for storing data in zlog or in ram (using std::vector) for development. However, we are adding a ram-based backend for zlog which means that we can remove the backend abstraction for the key-value store and use the zlog backend exclusively.

This issue would remove the key-value store backend abstraction and explicitly use zlog as the only choice.

The kvstore implementation currently handles the most basic case: serial intention.

With cmake the ceph backend will be built automatically if the correct dependencies exist. The option can be explicitly turned off using -DWITH_CEPH_BACKEND=0. Specifying -DWITH_CEPH_BACKEND=1 is the same as the default when no option is specified.

This issue would add a third mode that distinguishes between the default ON state and the option being explicitly requested. In the later case we want to fail/abort configuration when the dependencies are not found.

See CMakeLists.txt in the root of the project for the current set of options and dependency checks.

The binary tree used to represent the key-value store database is traversed in a few different ways. For instance there is a set of methods that traverse the tree to create graphviz/dot output of the tree structure. Another traversal is used for debugging which asserts that a tree is a valid red-black tree. It would be nice to add a few traversal types that can be used to plugin visitors to avoid duplicating the traversal code.