Scalaris, a distributed transactional key-value store
Scalaris is a scalable, transactional, distributed key-value store. It was the first NoSQL database that supported the ACID properties for multi-key transactions. It can be used for building scalable Web 2.0 services.Scalaris uses a structured overlay with a non-blocking Paxos commit protocol for transaction processing with strong consistency over replicas. Scalaris is implemented in Erlang.
Documentation / Download / Discussion:
- Users and Developers Guide (download as pdf)
- FAQ
- Prebuild packages
- Mailing list
Current Stable Release
Scalaris 0.6.0 (codename "Conus scalaris") - August 19, 2013
Packaging
- add ArchLinux packages
- add support for new distribution versions
API
- no more timeouts in client APIs
- Java-API: re-worked the request and result list handling -> move result processing to the operation classes
- Java-API: better support for custom operations
- Java-API: support the new partial reads: ReadRandomFromListOp and ReadSublistOp
- Java-API: compile with "vars" debug info
- Java-API: integrate new OtpErlang library (1.5.8 from R16B) with fixed support for compressed binaries
- Java-API: add back-ports from the Wiki on Scalaris demonstrator:
- list-change operations: ScalarisChangeListOp and ScalarisListAppendRemoveOp
- MultiMap classes are now in de.zib.tools
- CircularByteArrayOutputStream
- Java-API: fix hostname issues with Erlang and Java
- Java-API: slightly changed the delete API
- JSON-API: add API for auto-scale requests
- Python-API: add API for auto-scale requests
- Python-API: use default socket timeout
- Ruby-API: use default socket timeout
- all APIs: support lists of composite types
Demonstrator "Wiki on Scalaris"
(supported by 4CaaSt http://www.4caast.eu/ and Contrail http://contrail-project.eu):- allow monitoring via JMX in the FourCaastMonitoringPlugin
- support for getting random articles via the new partial read op
- new optimisation scheme "Buckets with Write Cache" - uses a single big list
- to read from and the rest of the buckets to write to
- improve import and dump-processing (faster, more memory-efficient)
- add on-the-fly conversion to the different optimisation schemes during import
- (only one prepared DB dump needed now)
- several UI enhancements and rendering fixes
- update bliki lib (includes code ported to upstream)
- add auto-import ability
- use tomcat 7.0.33
Business Logic
- replace common message tags with integers to reduce bandwidth
- more flexible read operations (easier to extend)
- add support for the following partial reads: random_from_list and sublist
- save bandwidth by not returning the full value for write operations (only the version is required)
- new DB back-end implementation with a smaller and cleaner interface
- faster DB get_chunk processing
- tx: allow overwriting old/outdated DB entries
- tx: allow overwriting old/outdated write-locked entries
- tx: allow setting write lock on old/outdated read-locked entries
- tx: always reply when the majority replied during read
- tx: make sure that if not_found is reported to the user (while reading), a write cannot go through if it is not also based on not_found
- tx: committing a test_and_set op on a non-existing entry now fails as well (the op itself already returned the failure)
- tx: add a 2s delay to wait for slow learner_decide answers before cleaning up (results in a faster state cleanup after the fourth response)
- tx: small performance improvements in several modules
- rm: only add alive, non-leaving nodes
- rm: if a predecessor crashes, start repairing the range (rrepair)
- rrepair: stabilised rrepair (not considered experimental any more)
- rrepair: also update entries with existing but outdated WriteLocks
- rrepair: several performance improvements (bloom, merkle_tree, art and rrepair processes in general)
- rrepair: re-design of rr_recon
- rrepair: don't offload heavy work onto the dht_node (increases responsiveness of the dht_node process during replica repair)
- rrepair: improve db_generator tool and random_bias binomial distribution used for tests
- rrepair: support differently configured nodes (use the same reconciliation structure parameters)
- rrepair: de-activate self-repair (a node with multiple copies of the same items does not need a reconciliation structure to repair some of them)
- rrepair: activate rrepair periodically every 10 minutes with a probability of 33%
- slide v2.0: fewer message to initiate a slide
- slide v2.0: generic (asynchronous) call-backs for different ring maintenance algorithms
- slide v2.0: re-work handling of planned next operations, e.g. used by incremental slides
- slide v2.0: don't directly work on the DB any more (there may be more data needed to slide) - let dht_node_state decide
- slide v2.0: activate incremental join and leave operations
- slide v2.0: actively report graceful node shutdown to the local FD of the leaving node to inform subscribers
- slide v2.0: code clean-up
- slide v2.0: some fixes for incremental slides
- slide v2.0: more robust in general
- more smooth node joins by also reporting when a join is not possible due to a running slide at the existing node
- passive load balancing: random selection of (equally qualified) nodes
- add new routing algorithms FRT-Chord (flexible routing tables) and GFRT-Chord (supports proximity routing and data centers) as alternatives to Chord (see rt_frtchord and rt_gfrtchord modules)
- add auto-scale framework, e.g. for cloud environments (supported by Contrail http://contrail-project.eu/) which is able to scale the deployment to maintain a given target latency of executed transactions
- cache config reads in the process dictionary for better performance
- cyclon: if the cache is empty, try one of the nodes in known_hosts
- add support for consistent snapshots (experimental)
Infrastructure
- add a daemon to monitor Scalaris via JMX
- disable message compression (only client values are compressed - the rest is too expensive, at least on GbE)
- support for distributions with python3 available as "python" and python2 as "python2"
- support for Ruby 1.9
- yaws 1.96 (with patch to compile on otp master and a patch to fix a performance regression)
- support for Erlang R13B01 up to R16B01 and current otp master
Tests
- add test suite to find memory leaks
- let "make test" run the major test suites and "make test-skipped" for some more (time-consuming) tests
- clean-up ring after timetrap timeout failures via common test hook
- new ?compare macro for custom comparison functions
- higher test coverage with more random-testing via the "tester"
Documentation
- user-dev-guide: add user tutorial on using scalaris
- user-dev-guide: add a section about the slide protocol
- user-dev-guide: extended description of scientific background
- add replica repair sequence diagrams
- better code descriptions
Tools
- gen_component: synchronous breakpoint set and delete for more deterministic usage
- trace_mpath: allow selective tracing via filter fun
- trace_mpath: fix several triggers becoming infected by trace_mpath resulting in infinite tracing
- trace_mpath: improve latex output of traces
- tester: copy dictionary to worker threads
- tester: add support for more types, e.g. neg_integer(), gb_rees
- tester: better type check error reporting
- tester: print tester last calls when aborting unit tests (timeout or exception)
- tester: add support for constraints in type specs ("when is_subtype(A,B)")
- web debug interface: add cluster graph visualisation
- web debug interface: display vivaldi distance
- web debug interface: add IP addresses and ports to the ring charts and tables
- web debug interface: allow navigating to the web interfaces of shown nodes
- top: support for showing messages in message queue of an inspected PID
- top: support for showing larger dictionary values
- allow recursive reply_as envelopes
- experimental protocol scheduler to check protocols with random message interleavings (see proto_sched module)
Bugs
- fix RM handling of (out-dated) nodes with the same ID as newly added nodes
- fix ganglia integration not working any more
- restore the ability to start nodes at a specific key via scalarisctl -k <key> ...
- fix some memory leaks in the tx system
- fix statistics of comm_connection (not send in some cases, not overflow-aware)
- use /bin/bash instead of /bin/sh which may not result in a bash session
- fix init.d scripts not checking for existing processes correctly
- fix dc_clustering
- fix numerous other bugs
