Apache Cassandra client and transport architecture. Slides from ApacheCon NA 2013 in Portland, OR.

1. Apache Cassandra
Clients and
Transports
Thursday, February 28, 13

2. Hi Folks!
I’m Nate
@zznate
Thursday, February 28, 13

3. API Management
API Analytics
API Tools
Thursday, February 28, 13

4. Clients and transports
for Cassandra
Thursday, February 28, 13

5. But first...
some questions
Thursday, February 28, 13

6. But first:
Architectural stuff
Thursday, February 28, 13

7. Cassandra:
“Sparsely Columnar”
Thursday, February 28, 13

8. An RDBMS table
Thursday, February 28, 13

9. An RDBMS table
Thursday, February 28, 13

10. Cassandra Style
Thursday, February 28, 13

11. Cassandra Style
Thursday, February 28, 13

12. Cassandra data
modelling
Thursday, February 28, 13

13. Four common patterns
Simple object to simple row
Sparse object to rows
Materialized view
Manual index
Thursday, February 28, 13

14. simple objects to
simple row
Thursday, February 28, 13

15. “static” column family
Thursday, February 28, 13

16. Sparse Objects
Thursday, February 28, 13

17. “dynamic” column family
Thursday, February 28, 13

18. Materialized Views
Thursday, February 28, 13

19. Materialized view
Thursday, February 28, 13

20. Regardless of the
approached used, there
are four overall goals
Thursday, February 28, 13

21. 1. Denormalize
2. Eliminate seeks
3. Design for read
4. Optimiza for blind
writes
Thursday, February 28, 13

22. Now... let’s talk about
protocols
Thursday, February 28, 13

23. Thrift
Thursday, February 28, 13

24. Thrift
RPC-Based
Thursday, February 28, 13

25. Thrift
RPC-Based
Mature Apache Project
Thursday, February 28, 13

26. Thrift
RPC-Based
Mature Apache Project
Supports lots of languages
Thursday, February 28, 13

27. Thrift
RPC-Based
Mature Apache Project
Supports lots of languages
Extensible!
Thursday, February 28, 13

28. CQL
Thursday, February 28, 13

29. CQL
Well defined protocol
Thursday, February 28, 13

30. CQL
Well defined protocol
Supports Compression
Thursday, February 28, 13

31. CQL
Well defined protocol
Supports Compression
Netty/NIO-based
Thursday, February 28, 13

32. Storage Mechanics
(but quickly)
Thursday, February 28, 13

33. get_slice
Workhorse of Cassandra selection
methods
Thursday, February 28, 13

34. get_slice: key
The row key
Thursday, February 28, 13

35. get_slice: ColumnParent
The column family (a.k.a table)
Thursday, February 28, 13

36. get_slice: SlicePredicate
defines the column range, or
specifically named columns
Thursday, February 28, 13

37. get_slice:ConsistencyLevel
The level of consistency we want for
this read
Thursday, February 28, 13

38. Obtuse at first glance,
but nothing is hidden
Thursday, February 28, 13

39. So...
Thursday, February 28, 13

40. But one person’s
abstraction leakage is
another’s preffered
model
Thursday, February 28, 13

41. How closely do you
want to interact with
the underlying storage
engine?
Thursday, February 28, 13

42. Client APIs
Thursday, February 28, 13

43. Benefits of thrift
Thursday, February 28, 13

44. Benefits of thrift
Mature selection of clients
Thursday, February 28, 13

45. Benefits of thrift
Mature selection of clients
Multiple languages
Thursday, February 28, 13

46. Benefits of thrift
Mature selection of clients
Multiple languages
Well documented
Thursday, February 28, 13

47. Benefits of thrift
Mature selection of clients
Multiple languages
Well documented
Can be used in other places
Thursday, February 28, 13

48. Drawbacks of thrift
Thursday, February 28, 13

49. Drawbacks of thrift
Several objects are required for
any request
Thursday, February 28, 13

50. Drawbacks of thrift
Several objects are required for
any request
Clients differs in implementation
Thursday, February 28, 13

51. Drawbacks of thrift
Several objects are required for
any request
Clients differs in implementation
Upstream dependency issues
Thursday, February 28, 13

52. Drawbacks of thrift
Several objects are required for
any request
Clients differs in implementation
Upstream dependency issues
Schema changes and cluster
health done pro-actively
Thursday, February 28, 13

53. Benefits of cql api
Thursday, February 28, 13

54. Benefits of cql api
Stored procedures
Thursday, February 28, 13

55. Benefits of cql api
Stored procedures
Common operations are
straight forward
Thursday, February 28, 13

56. Benefits of cql api
Stored procedures
Common operations are
straight forward
Cluster health and schema
change push-back
Thursday, February 28, 13

57. Benefits of cql api
Stored procedures
Common operations are
straight forward
Cluster health and schema
change push-back
Awesome client available
Thursday, February 28, 13

58. Drawbacks of CQL apis
Thursday, February 28, 13

59. Drawbacks of CQL apis
Still have idiomatic clients
Thursday, February 28, 13

60. Drawbacks of CQL apis
Still have idiomatic clients
Still a binary protocol
Thursday, February 28, 13

61. Drawbacks of CQL apis
Still have idiomatic clients
Still a binary protocol
Default storage model
emposes substantial
restrictions
** see gotchas section later
Thursday, February 28, 13

62. Considerations for your
app
Thursday, February 28, 13

63. Stick with Thrift if...
Thursday, February 28, 13

64. Heavy update
workloads
Thursday, February 28, 13

65. Large, dynamic batch
insertions
Thursday, February 28, 13

66. Hadoop integration
(CASSANDRA-4421)
Thursday, February 28, 13

67. Commonly deal with
very wide rows
(CASSANDRA-4176)
Thursday, February 28, 13

68. CASSANDRA-4176:
“Pick your shard keys
carefully”
Thursday, February 28, 13

69. Thursday, February 28, 13

70. Consider CQL if...
Thursday, February 28, 13

71. Static column family
model:
Take advantage of
stored procedures for
common reads
Thursday, February 28, 13

72. Despite the shard key
jab, CQL makes good
use of the storage
model
Thursday, February 28, 13

73. You can replace some
custom serialization
with collections
Thursday, February 28, 13

74. Integration with JDBC
and/or BI tools
Thursday, February 28, 13

75. Wire efficient:
Does not return
timestamp or TTL by
default
Thursday, February 28, 13

76. Larger, potentially more
transient evironments
Thursday, February 28, 13

77. But CQL is
- new
- an abstraction
Thursday, February 28, 13

78. In some cases, CQL
might not do what you
think
Thursday, February 28, 13

79. Most common CQL
pitfalls
Thursday, February 28, 13

80. Collections can only
be retrieved in their
entirety
Thursday, February 28, 13

81. Can’t mix static and
dynamic data in a
column family
Thursday, February 28, 13

82. “keys only” range
slices don’t work
(CASSANDRA-4536)
Thursday, February 28, 13

83. Range ghosts will not
be returned
Thursday, February 28, 13

84. Batch inserts are
clunky
(CASSANDRA-4693)
Thursday, February 28, 13

85. With non-compact
storage the whole row
must be read every
time.
Thursday, February 28, 13

86. The take away is that
you have options.
Particularly good ones
for Java.
Thursday, February 28, 13

87. Thursday, February 28, 13

88. BUT
Thursday, February 28, 13

89. there is a larger, more
fundamental problem
to discuss
Thursday, February 28, 13

90. “If [they] think that CQL is the
answer to usability then I just
won. We at least know where
our problems are.”
- 10gen exec.
Thursday, February 28, 13

91. The market has spoken
and we missed the
boat.
Thursday, February 28, 13

92. POST /endpoint {json}
Thursday, February 28, 13

93. A Cassandra-MVP
actually maintains a
REST front-end
Thursday, February 28, 13

94. So we’ve taken this
and gone further
Thursday, February 28, 13

95. What if...
Thursday, February 28, 13

96. Coming soon...
Intravert.
Vert.x+Cassandra.
ASF-licensed.
Driven by real-world
requirements.
Thursday, February 28, 13