Skip to content

Latest commit

 

History

History
153 lines (114 loc) · 5.55 KB

CHANGELOG.mdown

File metadata and controls

153 lines (114 loc) · 5.55 KB
Raw

0.7.5

  • Bug fix for potential SIGSEGV caused by invalid lock on filter check
  • Fixed circular buffer writing bug
  • Git SHA: 16a99fd

0.7.4

  • Bug fix for potential issues related to non-ASCII filter names
  • Git SHA: 619ab48

0.7.3

  • Adding support for a bind_address
  • Git SHA: a0706a9

0.7.2

  • Fixing issue with flush and unmap being invoked too often due to integer divide bug
  • Git SHA: 455d38b

0.7.1

  • Increase the safety of connection handling. Prevents some crashing conditions.
  • Reduce checkpoint intervals from 1 second to 250 milliseconds. This allows faster cleanups, and helps prevent "Delete in progress" errors.
  • Git SHA: 5715bea

0.7.0

  • Major rewrite of the internal Multi-Version Concurrency Controll (MVCC). The old system was based on a single map per version, along side a vacuum thread to do periodic cleanup. This did not scale well as each version had to copy the map of the last version, and since many versions could exist concurrently, it is possible for the heap to explode to many hundreds of MB or GB worth of just versioning information. Instead, we now maintain only 2 maps, a "primary" and "alternate", and make use of a list of delta updates. The vacuum thread now manages applying delta updates to the alternate map and rotating the primary and alternate. This reduces the memory overhead, and reduces the cost of creating a new version (create, drop, clear).
  • List ordering is no defined due to delta lists
  • Slow flush and unmap operations still allow vacuum to make progress
  • Git SHA: 8dc38de

0.6.6

  • Bug fixes for filters with long prefixes
  • Git SHA: dbce6b2

0.6.5

  • Replace internal hashmap with an Adaptive Radix Tree (ART). This improves the create/drop time for instances with many many filters.
  • list command now can take a prefix and returns only filters matching
  • As a result of the ART tree, list results are sorted in descending order
  • Git SHA: 238e85f

0.6.1 (unreleased)

  • Fixing thread safety issue with new-style memory buffers that could potentially lead to either pages not being flushed, or pages with no changes being re-flushed.

0.6.0

  • Optimize creation of new bitmaps. Dramatic speedup for create on new filters, or doing a resize.
  • Bloomd will no longer terminate if it fails to create a filter, instead an internal error will be reported.
  • More robust checkpointing mechanism to improve the safely of the vacuum thread.
  • create command can now return "Delete in progress" error. This is to expose the fact that a previously issued delete or clear may not yet be complete, and that the client should retry the create. Previously, the system could run into internal consistency issues.
  • Git SHA: 8f045fc

0.5.1

  • Added error message if the magic byte is incorrect
  • Perform an early flush when filters are created to ensure the config.ini file and bitmap headers are written
  • Fixed possible crashing bug
  • Fixed possible race condition in vacuum thread
  • Fixing build issues on OSX 10.8
  • Fixing validation of filter names
  • Added link to RPM repo, thanks to @penpen
  • Many bug fixes and improvements, thanks to @cemeyer
  • Git SHA: da8d3c8

0.5.0

  • Adding in_memory flag for the info command. This avoids needing to use a heuristic based on page_in / page_outs to determine the status of a filter.

  • Filter manager uses a simple form of Multi-Version Concurrency Control (MVCC) to enable progress to be made on operations in the face of potentially slow management operations. Increases read concurrency within the filter manager.

  • Changed networking from a leader/follower model to fixing connections to threads. This reduces locking contention in the networking layer and enables much higher concurrency in the workers.

  • Adding new memory buffer management that gives bloomd more control as opposed to relying on the kernel to flush buffers. This can now be controlled using a new use_mmap flag. If this flag is not provided, it defaults to 0, which instructs bloomd to use anonymous buffers and to manually manage flushing to disk. Otherwise, if it is 1, then files are mapped in using MAP_SHARED, and msync() is used for flushing. The advatage of having bloomd manage buffers is that under high I/O and memory pressure, the kernel will begin to block writes to mapped memory until they can be paged to disk which causes availability issues. With our own buffers, the flush thread will instead become increasingly behind, but the server can continue to serve requests. The down side is, if bloomd were to crash, we will take data loss as the buffers have not yet been written out. With MAP_SHARED buffers, the kernel will take care to persist the data.

  • Git SHA: 6dfd4b5

0.1.5

  • Several GCC compatibility changes to avoid unused result warnings on 4.7
  • Bug fix that affects newer glibc versions regarding usage of writev()
  • Fixed compatibility with Ubuntu 12.04
  • Git SHA: 76dbc38

0.1.4

  • Added new clear command
  • Git SHA: a904123

0.1.3

  • Fixed bug that would be encountered when switching to buffered output that caused endless buffer resizing.
  • Git SHA: 649c804

0.1.2

  • Flush command does not acquire read lock, remove the Disk I/O bottleneck when flush is happening.
  • Git SHA: f621be8

0.1.1

  • Fixed a bug with faulting in filters with multiple underlying data files that would cause resizing to take place unnecessarily.
  • Git SHA: c6c3b32

0.1.0

  • Initial Release
  • Git SHA: 718a438