Stephan Ewen - Experiences running Flink at Very Large Scale

Experiences Running
Apache Flink at
Very Large Scale
@StephanEwen
Berlin Buzzwords, 2017
1

3
 Various use cases
• Example: Stream ingestion, route events to Kafka, ES, Hive
• Example: Model user interaction sessions
 Mix of stateless / moderate state / large state
 Stream Processing as a Service
• Launching, monitoring, scaling, updating
@

5
 Blink based on Flink
 A core system in Alibaba Search
• Machine learning, search, recommendations
• A/B testing of search algorithms
• Online feature updates to boost conversion rate
 Alibaba is a major contributor to Flink
 Contributing many changes back to open source
@

7
@
Social network implemented using event sourcing
and CQRS (Command Query Responsibility
Segregation) on Kafka/Flink/Elasticsearch/Redis
More: https://data-artisans.com/blog/drivetribe-cqrs-apache-flink

How we learned to view Flink
through its users
8

System for Event–driven Applications
9
Event-driven
Applications
Stream Processing
Batch Processing
Stateful, event-driven,
event-time-aware processing
(event sourcing, CQRS, …)
(streams, windows, …)
(data sets)

Event Sourcing + Memory Image
10
event log
persists events
(temporarily)
event /
command
Process
main memory
update local
variables/structures
periodically snapshot
the memory

Event Sourcing + Memory Image
11
Recovery: Restore snapshot and replay events
since snapshot
event log
persists events
(temporarily)
Process

Distributed Memory Image
12
Distributed application, many memory images.
Snapshots are all consistent together.

Stateful Event & Stream Processing
13
Scalable embedded state
Access at memory speed &
scales with parallel operators

14
Re-load state
Reset positions
in input streams
Rolling back computation
Re-processing

15
Restore to different
programs
Bugfixes, Upgrades, A/B testing, etc

Compute, State, and Storage
16
Classic tiered architecture Streaming architecture
database
layer
compute
layer
application state
+ backup
compute
+
stream storage
and
snapshot storage
(backup)
application state

System for Event–driven Applications
17
Event-driven
Applications
Stream Processing
Batch Processing
Stateful, event-driven,
event-time-aware processing
(event sourcing, CQRS, …)
(streams, windows, …)
(data sets)

Apache Flink's Layered APIs
18
Process Function (events, state, time)
DataStream API (streams, windows)
Table API (dynamic tables)
Stream SQL
Stream- &
Batch Processing
Analytics
Stateful
Event-Driven
Applications

Lessons Learned from Running
Flink
19

20
The event/stream pipeline
generally just works


Interacting with the environment
 Dependency conflicts are amongst the biggest problems
• Next versions trying to radically reduce dependencies
• Make Hadoop an optional dependency
• Rework shading techniques
 The deployment ecosystem is crazy complex
• Yarn, Mesos & DC/OS, Docker & K8s, standalone, …
• Containers and overlay networks are tricky
• Authorization and authentication ecosystem complex it itself
• Continuous work to improve integration
21

External systems
 Dependency on any external system eventually causes
downtime
• Mainly: HDFS / S3 / NFS / … for checkpoints
 We plan to reduce dependency on those more and more in
the next versions
22

Type Serialization
 Type serialization is a harder problem in streaming than in
batch
• The data structure updates require more serialization
• Types are often more complex than in batch
 State lives long and across jobs
• Requires to "version" state and serializers
• Requires a "schema evolution" path
• Much enhanced support in Flink 1.3, more still to come
23

24
…is the most important part of
running a large scale Flink application
Robustly checkpointing…

Review: Checkpoints
25
Trigger checkpoint Inject checkpoint barrier
stateful
operation
source /
transform

Review: Checkpoints
26
Take state snapshot Trigger state
snapshot
stateful
operation
source /
transform

Review: Checkpoint Alignment
27
begin aligning
checkpoint
barrier n
xy
operator
aligning
ab
operator
23 1
input buffer
y

Review: Checkpoint Alignment
28
bc
operator
23 1
emit barrier n
c
operator
23 1
input buffer
continuecheckpoint
4 4
a

Understanding Checkpoints
30
How well behaves
the alignment?
(lower is better)
How long do
snapshots take?
delay =
end_to_end – sync – async

Understanding Checkpoints
31
How well behaves
the alignment?
(lower is better)
How long do
snapshots take?
delay =
end_to_end – sync – async
long delay = under backpressure
under constant backpressure
means the application is
under provisioned
too long means
 too much state
per node
 snapshot store cannot
keep up with load
(low bandwidth)
vastly improved with
incremental checkpoints in Flink 1.3
most important
robustness metric

Heavy alignments
 A heavy alignment typically happens at some point
 Different load on different paths
 Skewed window emission
(lots of data on one node)
 Stall of one operator on the path
34

Heavy alignments
35

Heavy alignments
36
GC stall

Catching up from heavy alignments
 Operators that did heavy alignment need to catch up again
 Otherwise, next checkpoint will have a
heavy alignment as well
37
operator
bc
operator
23 14
a
consumed first after
checkpoint completed
bc a

Catching up from heavy alignments
 Giving the computation time to catch up before starting the
next checkpoint
• Set the min-time-between-checkpoints
• Ideas to change checkpoints to policy based (spend x% of capacity
on checkpoints)
 Asynchronous checkpoints mitigate most of problem
• Very short stalls in the pipelines means shorter alignment phase
• Catch up already happens concurrently to state materialization
38

Asynchrony of different state types
40
State Flink 1.2 Flink 1.3 Flink 1.4
Keyed state
RocksDB ✔ ✔ ✔
Keyed State
on heap
✘ (✔)
(hidden in 1.2.1) ✔ ✔
Timers ✘ ✘ ✔ (PR)
Operator State ✘ ✔ ✔

When to use which state backend?
41
Async. Heap/FS RocksDB
State ≥ Memory ?
Complex Objects?
(expensive serialization)
high data rate?
no yes
yes no
yes
no
a bit
simplified

We are hiring!
data-artisans.com/careers

Avoiding DDOSing other systems
45

Exceeding FS request capacity
 Job size: multiple 1000 operators
 Checkpoint interval: few secs
 State size: KBs per operator, 1000 of state chunks
 Via the S3 FS (from Hadoop), writes ensure "directory"
exists, 2 HEAD requests
 Symptom: S3 blocked off connections after exceeding
1000s HEAD requests / sec
46

Reducing FS stress for small state
47
JobManager TaskManager
Checkpoint
Coordinator
Task
TaskManager
Task
TaskTask
Root Checkpoint File
(metadata) checkpoint data
files
Fs/RocksDB state backend
for most states

Reducing FS stress for small state
48
Checkpoint
Coordinator
Task
TaskManager
Task
TaskTask
checkpoint data
directly in metadata file
Fs/RocksDB state backend
for small states
ack+data
Increasing small state
threshold reduces number
of files (default: 1KB)

Deploying Tasks
50
Happens during initial deployment and recovery
Akka / RPC Akka / RPC
Blob Server Blob Server
Deployment RPC Call
Contains
- Job Configuration
- Task Code and Objects
- Recover State Handle
- Correlation IDs

Deploying Tasks
51
Happens during initial deployment and recovery
Blob Server Blob Server
Deployment RPC Call
Contains
- Job Configuration
- Task Code and Objects
- Recover State Handle
- Correlation IDs
KBs
up to MBs
KBs
few bytes

RPC volume during deployment
52
(back of the napkin calculation)
number of
tasks
2 MB
parallelism
size of task
objects
100010 x x
x x
=
= RPC volume
20 GB
~20 seconds on full 10 GBits/s net
> 1 min with avg. of 3 GBits/s net
> 3 min with avg. of 1GBs net

Timeouts and Failure detection
53
~20 seconds on full 10 GBits/s net
> 1 min with avg. of 3 GBits/s net
> 3 min with avg. of 1GBs net
Default RPC timeout: 10 secs
default settings lead to failed
deployments with RPC timeouts
Solution: Increase RPC timeout
Caveat: Increasing the timeout makes failure detection slower
Future: Reduce RPC load (next slides)

Dissecting the RPC messages
54
Message part Size
Variance across
subtasks
and redeploys
Job Configuration KBs constant
Task Code and Objects up to MBs constant
Recover State Handle KBs variable
Correlation IDs few bytes variable

Upcoming: Deploying Tasks
55
Out-of-band transfer and caching of
large and constant message parts
Blob Server Blob Cache
(1) Deployment RPC Call
(Recover State Handle,
Correlation IDs, BLOB pointers)
(2) Download and cache BLOBs
(Job Config, Task Objects) MBs
KBs

Layers of abstraction
56
Ogres have
layers
So do
squirrels

Apache Flink's Layered APIs
57
Process Function (events, state, time)
DataStream API (streams, windows)
Table API (dynamic tables)
Stream SQL
Stream- &
Batch Processing
Analytics
Stateful
Event-Driven
Applications

Process Function
58
class MyFunction extends ProcessFunction[MyEvent, Result] {
// declare state to use in the program
lazy val state: ValueState[CountWithTimestamp] = getRuntimeContext().getState(…)
def processElement(event: MyEvent, ctx: Context, out: Collector[Result]): Unit = {
// work with event and state
(event, state.value) match { … }
out.collect(…) // emit events
state.update(…) // modify state
// schedule a timer callback
ctx.timerService.registerEventTimeTimer(event.timestamp + 500)
}
def onTimer(timestamp: Long, ctx: OnTimerContext, out: Collector[Result]): Unit = {
// handle callback when event-/processing- time instant is reached
}
}

Data Stream API
59
val lines: DataStream[String] = env.addSource(
new FlinkKafkaConsumer09<>(…))
val events: DataStream[Event] = lines.map((line) => parse(line))
val stats: DataStream[Statistic] = stream
.keyBy("sensor")
.timeWindow(Time.seconds(5))
.sum(new MyAggregationFunction())
stats.addSink(new RollingSink(path))

Events, State, Time, and Snapshots
61

62
f(a,b)
Event-driven function
executed distributedly

63
f(a,b)
Maintain fault tolerant local state similar to
any normal application
Main memory +
out of core (for maps)

64
f(a,b)
wall clock
event time clock
Access and react to
notions of time and progress,
handle out-of-order events

65
f(a,b)
wall clock
event time clock
Snapshot point-in-time
view for recovery,
rollback, cloning,
versioning, etc.

66
Source
Transformation
Transformation
Sink
val lines: DataStream[String] = env.addSource(new FlinkKafkaConsumer09(…))
val events: DataStream[Event] = lines.map((line) => parse(line))
val stats: DataStream[Statistic] = stream
.keyBy("sensor")
.timeWindow(Time.seconds(5))
.sum(new MyAggregationFunction())
stats.addSink(new RollingSink(path))
Streaming
Dataflow
Source Transform Window
(state read/write)
Sink

67
Source
Filter /
Transform
State
read/write
Sink

68
Scalable embedded state
Access at memory speed &
scales with parallel operators

69
Re-load state
Reset positions
in input streams
Rolling back computation
Re-processing

70
Restore to different
programs
Bugfixes, Upgrades, A/B testing, etc

"Classical" versus
Streaming Architecture
71

Compute, State, and Storage
72
database
layer
compute
layer
application state
+ backup
compute
+
stream storage
and
snapshot storage
(backup)
application state

Performance
73
synchronous reads/writes
across tier boundary
asynchronous writes
of large blobs
all modifications
are local

Consistency
74
distributed transactions
at scale typically
at-most / at-least once
exactly once
per state
=1 =1snapshot consistency
across states

Scaling a Service
75
separately provision additional
database capacity
provision compute
and state together
provision compute

Rolling out a new Service
76
provision a new database
(or add capacity to an existing one)
provision compute
and state together
simply occupies some
additional backup space

Repair External State
77
Streaming architecture
events
live application external state
wrong results
backed up data
(HDFS, S3, etc.)

78
overwrite
with correct results
backed up data
(HDFS, S3, etc.)
application on backup input
events

79
overwrite
with correct results
backed up date
(HDFS, S3, etc.)
Each application doubles as
a batch job!
application on backup input
events

Stephan Ewen - Experiences running Flink at Very Large Scale

In this document

More Related Content

What's hot

Similar to Stephan Ewen - Experiences running Flink at Very Large Scale

More from Ververica

Recently uploaded

Stephan Ewen - Experiences running Flink at Very Large Scale