Skip to content

Latest commit

 

History

History
214 lines (156 loc) · 7.75 KB

CHANGELOG.md

File metadata and controls

214 lines (156 loc) · 7.75 KB

Table of Contents

2.4.0

Released on: 2019/03/28

Added

  • A new get_bulk() API allows for fetching several values from the layered caches in a single call, and will execute all L3 callback functions concurrently, in a configurable pool of threads. #77
  • purge() now clears the L1 LRU cache with the new flush_all() method when used in OpenResty >= 1.13.6.2. Thanks @Crack for the patch! #78

Fixed

  • get() is now resilient to L3 callback functions calling error() with non-string arguments. Such functions could result in a runtime error when LuaJIT is compiled with -DLUAJIT_ENABLE_LUA52COMPAT. Thanks @MartinAmps for the patch! #75
  • Instances using a custom L1 LRU cache in OpenResty < 1.13.6.2 are now restricted from calling purge(), since doing so would result in the LRU cache being overwritten. #79

Back to TOC

2.3.0

Released on: 2019/01/17

Added

  • Returning a negative ttl value from the L3 callback will now make the fetched data bypass the cache (it will still be returned by get()). This is useful when some fetched data indicates that it is not cacheable. Thanks @eaufavor for the patch! #68

Back to TOC

2.2.1

Released on: 2018/07/28

Fixed

  • When get() returns a value from L2 (shm) during its last millisecond of freshness, we do not erroneously cache the value in L1 (LRU) indefinitely anymore. Thanks @jdesgats and @javierguerragiraldez for the report and initial fix. #58
  • When get() returns a previously resurrected value from L2 (shm), we now correctly set the hit_lvl return value to 4, instead of 2. 307feca

Back to TOC

2.2.0

Released on: 2018/06/29

Added

  • Implement a new resurrect_ttl option. When specified, get() will behave in a more resilient way upon errors, and in particular callback errors. #52
  • New stale argument to peek(). When specified, peek() will return stale shm values. #52

Back to TOC

2.1.0

Released on: 2018/06/14

Added

  • Implement a new shm_locks option. This option receives the name of a lua_shared_dict, and, when specified, the mlcache instance will store lua-resty-lock objects in it instead of storing them in the cache hits lua_shared_dict. This can help reducing LRU churning in some workloads. #55
  • Provide stack traceback in err return value when the L3 callback throws an error. #56

Fixed

  • Ensure no memory errors returned by shm insertions are properly returned by set(). #53

Back to TOC

2.0.2

Released on: 2018/04/09

Fixed

  • Make get() lookup in shm after lock timeout. This prevents a possible (but rare) race condition under high load. Thanks to @jdesgats for the report and initial fix. #49

Back to TOC

2.0.1

Released on: 2018/03/27

Fixed

  • Ensure the set(), delete(), peek(), and purge() method properly support the new shm_miss option. #45

Back to TOC

2.0.0

Released on: 2018/03/18

This release implements numerous new features. The major version digit has been bumped to ensure that the changes to the interpretation of the callback return values (documented below) do not break any dependent application.

Added

  • Implement a new purge() method to clear all cached items in both the L1 and L2 caches. #34
  • Implement a new shm_miss option. This option receives the name of a lua_shared_dict, and when specified, will cache misses there instead of the instance's shm shared dict. This is particularly useful for certain types of workload where a large number of misses can be triggered and eventually evict too many cached values (hits) from the instance's shm. #42
  • Implement a new l1_serializer callback option. It allows the deserialization of data from L2 or L3 into arbitrary Lua data inside the LRU cache (L1). This includes userdata, cdata, functions, etc... Thanks to @jdesgats for the contribution. #29
  • Implement a new shm_set_tries option to retry shm:set() operations and ensure LRU eviction when caching values of disparate sizes. #41
  • The L3 callback can now return nil + err, which will be bubbled up to the caller of get(). Prior to this change, the second return value of callbacks was ignored, and users had to throw hard Lua errors from inside their callbacks. #35
  • Support for custom IPC module. #31

Fixed

  • In the event of a no memory error returned by the L2 lua_shared_dict cache (after the number of shm_set_tries failed), we do not interrupt the get() flow to return an error anymore. Instead, the retrieved value is now bubbled up for insertion in L1, and returned to the caller. A warning log is (by default) printed in the nginx error logs. #41

Back to TOC

1.0.1

Released on: 2017/08/26

Fixed

  • Do not rely on memory address of mlcache instance in invalidation events channel names. This ensures invalidation events are properly broadcasted to sibling instances in other workers. #27

Back to TOC

1.0.0

Released on: 2017/08/23

Initial release.

Back to TOC