The architecture of Akka Streams internally consists of several distinct layers:
* The DSLs like akka.stream.scaladsl.Flow, akka.stream.scaladsl.Source etc. are the user facing API
for composing streams. These DSLs are a thin wrappers around the internal akka.stream.impl.TraversalBuilder
builder classes. There are Java alternatives of these DSLs in javadsl which basically wrap their scala
counterpart, delegating method calls.
* The akka.stream.stage.GraphStage API is the user facing API for creating new stream operators. These
classes are used by the akka.stream.impl.fusing.GraphInterpreter which executes islands (subgraphs) of these
* The high level DSLs use the akka.stream.impl.TraversalBuilder classes to build instances of
akka.stream.impl.Traversal which are the representation of a materializable stream description. These builders
are immutable and safely shareable. Unlike the top-level DSLs, these are classic, i.e. elements are treated as
* The akka.stream.impl.Traversal is the immutable, efficient representation of a stream processing graph that
can be materialized. The builders exists solely for the purpose of producing a traversal in the end.
* The akka.stream.impl.PhasedFusingActorMaterializer is the class that is responsible for traversing and
interpreting a akka.stream.impl.Traversal. It delegates the actual task of creating executable entities
and Publishers/Producers to akka.stream.impl.PhaseIslands which are plugins that understand atomic operators
in the graph and able to turn them into executable entities.
* The akka.stream.impl.fusing.GraphInterpreter and its actor backed wrapper akka.stream.impl.fusing.ActorGraphInterpreter
are used to execute synchronous islands (subgraphs) of akka.stream.stage.GraphStages.
For the execution layer, refer to akka.stream.impl.fusing.GraphInterpreter.
The central piece for both the DSLs and materialization is the akka.stream.impl.Traversal. This is the
representation of an Akka Stream, basically a akka.stream.scaladsl.RunnableGraph. The design goals
for akka.stream.impl.Traversal are:
* Be able to materialize a graph in one pass over the traversal
* Unify materialization and fusing. The materializer should be able to construct all the necessary data structures
for the interpreters and for connecting them in one go.
* Avoid allocations as much as possible.
* Biased implementation for the 90% case. Common cases should be as fast as possible:
* wiring linear chains should be very fast.
* assume that most graphs are mostly linear, with only a few generalized graph constructs thrown in.
* materialization should not pay the price of island tracking if there is only a single island
* assume that the number of islands is low in general
* avoid "copiedModule" i.e. wrappers that exist solely for the purpose of establishing new port identities
for operators that are used multiple times in the same graph.
* Avoid hashmaps and prefer direct array lookup wherever possible
Semantically, a traversal is a list of commands that the materializer must execute to turn the description to a
running stream. In fact, the traversal is nothing more than an immutable list, that is expressed as a tree. A
tree is used to make immutable appends fast (immutable lists only have prepend as O(1) operation, append is O(N)).
The materializer "recovers" the original sequence by using a local, mutable stack to properly traverse the tree
structure. This is way cheaper than to immutably append to the traversal at each addition.
The "tree-ness" is expressed by explicit akka.stream.impl.Concat nodes that express that two traversals
need to be traversed in a certain sequence, stashing away the second on a local stack until the first is fully
While traversing the traversal (basically following Concat nodes), the materializer will visit the following
* akka.stream.impl.MaterializeAtomic: An atomic module needs to be materialized. This node also contains
wiring information which we discuss later.
* Materialized value computation. This is a stack based "sublanguage" to compute the final materialized value
on a stack, maintained by the materializer
* akka.stream.impl.PushNotUsed push a NotUsed value on the stack
* akka.stream.impl.Pop pop the top of the stack and throw away
* akka.stream.impl.Transform take the top of the stack, transform it with the provided function and put
the result back on the top of the stack
* akka.stream.impl.Compose take the top two values of the stack, invoke the provided function with these
values as arguments, then put the calculated value on the top of the stack
* Materialized values of atomic operators when visiting a akka.stream.impl.MaterializeAtomic must be
pushed to the stack automatically. There are no explicit PUSH commands for this
* Attributes calculation. These also are a stack language, although much simpler than the materialized value
commands. For any materialized operator, the top of the attributes stack should be provided as the current
* akka.stream.impl.PushAttributes combines the attributes on the top of the stack with the given ones and
puts the result on the attributes stack
* akka.stream.impl.PopAttributes removes the top of the attributes stack.
* Island tracking. Islands serve two purposes. First, they allow a large graph to be cut into parts that execute
concurrently with each other, using asynchronous message passing between themselves. Second, they are an
extension point where "plugins" (akka.stream.impl.PhaseIsland) can be used to specially handle subgraphs.
Islands can be nested in each other. This makes "holes" in the parent island. Islands also need a stack
as exiting a "hole" means returning to the parent, enclosing island and continuing where left.
* akka.stream.impl.EnterIsland instructs the materializer that the following commands will belong to
the materialization of a new island (a subgraph). The akka.stream.impl.IslandTag signals to
the materializer which akka.stream.impl.PhaseIsland should be used to turn operators of this island into
* akka.stream.impl.ExitIsland instructs the materializer that the current island is done and the parent
island is now the active one again.
Please note that the stack based materialized value computation eliminates the issues present in the older
materializer which expressed these computations as an AST. We had to use optimizations for this tree so that
long Keep.left chains don't explode the stack visiting a large AST. The stack based language sidesteps this issue
completely as the number of these commands don't increase the stack space required to execute them, unless the
computation itself requires it (which is not the case in any sane stream combination).
As a mental model, the wiring part of the Traversal (i.e. excluding the stack based sub-commands tracking
materialized values, attributes, islands, i.e. things that don't contribute to the wiring structure of the graph)
translates everything to a single, global, contiguous Array. Every input and output port of each operator is mapped
to exactly one slot of this "mental array". Input and output ports that are considered wired together simply map
to the same slot. (In practice, these slots might not be mapped to an actual global array, but multiple local arrays
using some translation logic, but we will explain this later)
Input ports are mapped simply to contiguous numbers in the order they are visited. Take for example a simple
Operator1[in1, in2, out] - Operator2[out] - Operator3[in]
This results in the following slot assignments:
* Operator1.in1 -> 0
* Operator1.in2 -> 1
* Operator3.in -> 2
The materializer simply visits Stage1, Stage2, Stage3 in order, visiting the input ports of each operator in order.
It then simply assigns numbers from a counter that is incremented after visiting an input port.
(Please note that all akka.stream.impl.StreamLayout.AtomicModules maintain a stable order of their ports, so
this global ordering is well defined)
Before explaining how output wiring works, it is important to settle some terminology. When we talk about ports
we refer to their location in the "mental array" as slots. However, there are other entities that needs to
reference various positions in this "mental array", but in these cases we use the term _offset_ to signify that
these are only used for bookkeeping, they have no "place" in the "array" themselves. In particular:
* offset of a module: The offset of an akka.stream.impl.StreamLayout.AtomicModule is defined as the value of
the input port counter when visiting the akka.stream.impl.MaterializeAtomic node to materialize that module.
In other words, the offset of a module is the slot of its first input port (if there is any). Since modules
might not have any input ports it can be that different modules share the same offset, simply because the
the first one visited does not increase the input port counter.
* offset of segments, islands: Defined similarly to module. The offset of an island or a segment is simply the
value of the input port counter (or the first unallocated slot).
Module1[in1 = 0, in2 = 1] - Module2[out] - Module3[in = 2]
The offset of Module1 is 0, while Module2 and Module3 share the same offset of 2. Note that only input ports (slots)
contribute to the offset of a module in a traversal.
Output ports are wired relative to the offset of the module they are contained in. When the materializer visits
a akka.stream.impl.MaterializeAtomic node, it contains an Array that maps ports to a relative offset. To
calculate the slot that is assigned to an output port the following formula is used:
slot = offsetOfModule + outToSlots(out.id)
Where outToSlots is the array contained in the akka.stream.impl.MaterializeAtomic node.
The power of this structure comes from the fact that slots are assigned in a relative manner:
* input ports are assigned in sequence so the slots assigned to the input ports of a subgraph depend on the
subgraph's position in the traversal
* output ports are assigned relative to the offset of their owner module, which is in turn relative to its first
(potential) input port (which is relative, too, because of the previous point)
This setup allows combining subgraphs without touching their internal wirings as all their internal wirings will
properly resolve due to everything being relative:
| | | |
|---------Graph1---------|--- .... ---|----Graph2----|
It is important to note that due to reusability, an Akka Stream graph may contain the same atomic or composite
multiple times in the same graph. Since these must be distinguished from each other somehow, they need port mapping
(i.e. a new set of ports) to ensure that the ports of one graph are distinguishable from another. Because how
the traversal relative addressing works, these are _temporary_ though, once all internal wirings are ready, these
mappings can be effectively dropped as the global slot assignments uniquely identify what is wired to what. For
example since Graph1 is visited before Graph2 all of the slots or offsets it uses are different from Graph2 leaving
no room for misinterpretation.
Port mapping is the way how the DSL can distinguish between multiple instances of the same graph included multiple
times. For example in the Graph DSL:
val merge1 = builder.add(Merge)
val merge2 = builder.add(Merge)
the port merge1.out must be different from merge2.out.
For efficiency reasons, the linear and graph DSLs use different akka.stream.impl.TraversalBuilder types to
build the akka.stream.impl.Traversal (we will discuss these next). One of the differences between the two
builders are their approach to port mapping.
The simpler case is the akka.stream.impl.LinearTraversalBuilder. This builder only allows building linear
chains of operators, hence, it can only have at most one OutPort and InPort unwired. Since there is no
possible ambiguity between these two port types, there is no need for port mapping for these. Conversely,
for those internal ports that are already wired, there is no need for port mapping as their relative wiring
is not ambiguous (see previous section). As a result, the akka.stream.impl.LinearTraversalBuilder does not
use any port mapping.
The generic graph builder class akka.stream.impl.CompositeTraversalBuilder needs port mapping as it allows
adding any kind of builders in any order. When adding a module (encoded as another akka.stream.impl.TraversalBuilder)
there are two entities in play:
* The module (builder) to be added. This builder has a few ports unwired which are usually packaged in a Shape
which is stored alongside with the builder in the Graph of the DSL. When invoking methods on this builder
these set of ports must be used.
* The module that we are growing. This module needs a new set of ports to be used as it might add this module
multiple times and needs to disambiguate these ports.
Adding to the akka.stream.impl.CompositeTraversalBuilder involves the following steps (pseudocode):
val newShape = shape.deepCopy() // Copy the shape of the module we want to add
val newBuilder = builder.add(submodule, newShape) // Add the module, and register it with the new shape
newBuilder.wire(newShape.in, ...) // Use the new ports to wire
What happens in the background is that Shape.deepCopy creates copies of the ports, and fills their
mappedTo field to point to their original port counterpart. Whenever we call wire in the outer module, it
delegates calls to the submodule, but using the original port (as the submodule builder has no knowledge of
the external mapping):
submodule.assign(port.mappedTo, ...) // enclosing module delegating to submodule, translating ports back
Visualizing this relationship:
| in', in" ---------+ | in' and in" both resolve to in
| | .mappedTo v .mappedTo | but they will be used on _different_ builders
| +-------------+ +-------------+ |
| | in | | in | | (delegation happens recursively in AddedModule)
| | AddedModule | | AddedModule | |
It is worth to note that the submodule might also continue this map-and-delegate chain to further submodules until
a builder is reached that can directly perform the operation. In other words, the depth of nesting is equal to
the length of mappedTo chaining.
IMPORTANT: When wiring in the enclosing module the new ports/shape MUST be used, using the original ports/shape
will lead to incorrect state.
In order to understand why builders are needed, consider wiring two ports together. Actually, we don't need to
wire input ports anywhere. Their slot is implicitly assigned by their position in the traversal, there is no
additional state we need to track. On the other hand, we cannot build a akka.stream.impl.MaterializeAtomic
node until the mapping array outToSlots is fully calculated. In other words, in reality, we don't wire input ports
anywhere, we only assign output ports to slots. The builders exist mainly to keep track all the necessary
information to be able to assign output ports, build the outToSlots array and finally the akka.stream.impl.MaterializeAtomic
node. The consequence of this that a akka.stream.impl.Traversal can be constructed as soon as all output ports
are wired ("unwired" inputs don't matter).
There is a specific builder that is used for the cases where all outputs have been wired:
akka.stream.impl.CompletedTraversalBuilder. This builder type simply contains the completed traversal plus
some additional data. The reason why this builder type exists is to keep auxiliary data structures required for
output port mapping only while they are needed, and shed them as soon as they are not needed anymore. Since builders
may recursively contain other builders, as soon as internals are completed those contained builders transition
to completed state and drop all additional data structures. This is very GC friendly as many intermediate graphs
exist only in a method call, and hence most of the additional data structures are dropped before method return and
can be efficiently collected by the GC.
The most generic builder is akka.stream.impl.CompositeTraversalBuilder. There are two main considerations
this builder needs to consider:
* Enclosed modules (builders) must have a stable position in the final traversal for relative addressing to work.
Since module offsets are calculated by traversal position, and outputs are wired relative to module offset, this
* Enclosed builders might not be complete yet (i.e. have unwired outputs) and hence they cannot immediately give
The composite builder keeps a temporary list of traversal steps (in reverse order because of immutable lists)
it needs to create once it is completed (all outputs wired). These steps refer to the traversal of submodules
as a akka.stream.impl.BuilderKey which is just a placeholder where the traversal of the submodule will be
stitched in. This akka.stream.impl.BuilderKey is also a key to a map which contains the evolving builder.
The importance of this "preimage" traversal is that it keeps position of submodules stable, making relative
Once the composite is completed, it takes these steps (now reversing it back to normal), and builds the traversal
using the submodule traversals referred to by akka.stream.impl.BuilderKey. Note that at this point all the
submodules are akka.stream.impl.CompletedTraversalBuilders because there are no unwired outputs and hence the
Traversal can be assembled. As the builder evolves over time, more and more of its akka.stream.impl.BuilderKeys
will refer to akka.stream.impl.CompletedTraversalBuilders, shedding much of the temporary data structures.
Refer to akka.stream.impl.CompositeTraversalBuilder for more details.
The akka.stream.impl.LinearTraversalBuilder is a much simpler beast. For efficiency, it tries to work as much
as possible directly on the akka.stream.impl.Traversal avoiding auxiliary structures. The two main considerations
for this builder are:
* akka.stream.scaladsl.Source and akka.stream.scaladsl.Flow contain an unwired output port. Yet, we would
like to build the traversal directly as much as possible, even though the builder is not yet completed
* akka.stream.impl.CompositeTraversalBuilders might be included in a linear chain. These cannot provide a
traversal before they are fully completed.
The linear builder, although it is one class, comes in basically two flavors:
* Purely linear builder: this contains only other linear builders, or all the composites that it includes
have been fully wired before and hence their traversal is now fully incorporated. Basically this kind
of builder only contains the akka.stream.impl.Traversal and only a couple of extra fields.
* Linear builder with an incomplete composite at the end (output): In this case, we have an incomplete composite.
It can only be at the end, since this is the only position where an output port can be unwired. We need to
carry this builder with us until the output port is finally wired, in which case we incorporate its traversal
into the already complete one, and hopefully transition to a purely linear builder.
If we consider the purely linear case, we still need to figure out how can we provide a traversal even though
the last output port is unwired. The trick that is used is to wire this output port optimistically to the
relative address -1 which is almost always correct (why -1? explained a bit later). If it turns out to be incorrect
later, we fix it by the helper method akka.stream.impl.Traversal.rewireFirstTo which tears down the traversal
until the wrong module is found, then fixes the port assignment. This is only possible on purely linear layouts though.
Again, this is an example of the 90% rule. Most appends will not need this rewiring and hence be as fast as possible
while the rarer cases suffering a minor penalty.
In the case where the last module is a composite, the above trick would not work as nothing guarantees that the
module that exposed its output port is at an expected position in the traversal. Instead, we simply keep around
this composite and delay construction of its part of the traversal. For details see akka.stream.impl.LinearTraversalBuilder
as these cases are heavily commented and explained in the code.
There is another peculiarity of the linear builder we need to explain. Namely, it builds the traversal in reverse
order, i.e. from Sinks towards Sources. THIS CAN BE SUPER CONFUSING AT TIMES SO PAY ATTENTION! There are two
important reasons why this is needed:
* Prepending to immutable lists is more efficient. Even though we encode our traversal list as a tree, we would
need stack space at materialization time as much as the length of the list if we would append to it instead of
* Prepending means that most output ports refer to slots visited before, i.e. output relative offsets are negative.
This means that during materialization, output ports will be wired to slots that the materializer visited before
which enables an efficient one-pass materialization design. The importance of this is discussed later below.
To visualize this, imagine a simple stream:
[Source.out] -> [Map.in, Map.out] -> [Sink.in]
offs = 0 offs = 1 offs = 1
[Sink.in = 0] <- [Map.in = 1, Map.out = -1] <- [Source.out = -1]
Since the traversal steps are reversed compared to the DSL order, it is important to reverse materialized value
All what we have discussed so far referred to the "mental array", the global address space in which slots
are assigned to ports. This model describes the wiring of the graph perfectly, but it does not map to the local
data structures needed by materialization when there are islands present. One of the important goals of this
layout data structure is to be able to produce the data structures used by the akka.stream.impl.fusing.GraphInterpreter
directly, without much translation. Unfortunately if there is an island inside a traversal, it might leave gaps
in the address space:
Since we visit Island2 before returning to Island1, the naive approach would leave a large gap between the last
input port visited before entering Island2 and the first input port visited when returning to Island1. What
we would like to have instead is a contiguous slot assignment from the viewpoint of Island1. This is where
akka.stream.impl.PhasedFusingActorMaterializer and its akka.stream.impl.IslandTracking helper comes into
the picture. These classes do the heavy-lifting of traversing the traversal and then mapping global slots to
slots local to the island, delegating then the local wiring to akka.stream.impl.PhaseIsland implementations.
For example the akka.stream.impl.GraphStageIsland sees only a contigous slot-space and hence it can directly
construct the array for the interpreter. It is not aware of the presence of other islands or how it is represented
in the global slot-space.
Materialzation is orchestrated by the akka.stream.impl.PhasedFusingActorMaterializer. It basically decodes the
traversal and handles islands. This top-level materializer does not really handle the wiring _inside_ an island,
it only handles wiring of Publishers and Subscribers that connect islands. Instead it delegates in-island wiring
to akka.stream.impl.PhaseIslands. For example a default fused island will be actually wired by
First, look at a traversal that has two islands:
In this traversal, we have two islands, and three, so called _segments_. Segments are simply contiguous range of
slots between akka.stream.impl.EnterIsland or akka.stream.impl.ExitIsland tags (in any combination). When
the materializer encounters either an enter or exit command, it saves various information about the segment it
just completed (what is its offset, how long it is measured in input slots, etc.). This information is later
used to figure out if a wiring crosses island boundaries or is it local to the island.
It is important to note that the data structure for this is only allocated when there are islands. This is again
the 90% rule in action. In addition, these data structures are java.util.ArrayList instances, where lookups
according to some value are implemented as simple linear scans. Since in 90% of the cases these structures are
very short, this is the most efficient approach. Cases where this can be a performance problem are very-very special
and likely not happen in practice (no graph should contain more than a dozen of islands for example).
When it comes to deciding whether a wiring is cross-island or local, there are two cases possible:
* we encountered an output port that is wired backwards (relative address is negative). In this case we already
have all the data necessary to resolve the question.
* we encountered an output port that is wired forward (relative address is positive). In this case we have not
yet visited that part of the traversal where the assignment points.
If we want to keep the one-pass design of the materializer, we need to delay forward wirings until we have all
the information needed, i.e. we visit the target in port. The akka.stream.impl.PhasedFusingActorMaterializer
has a data structure for tracking forward wires which it consults whenever it visits an input port. Again, this
is only allocated if needed, and it is again an array with linear scan lookup. Once the target input port have
been found, the rules of the wiring are the same as for backwards wiring.
backward wire (to the visited part) <------+ +------> forward wire (into the unknown)
|----Island1-----|----Island2(enclosed)-------- ... (this is where we are now)
Remember, the akka.stream.impl.LinearTraversalBuilder builds its akka.stream.impl.Traversal in backwards
order, so since most of the graphs are constructed by the linear DSLs almost all wirings will be backwards
(90% rule in action again).
When it comes to resolving wirings and calculating the local slots for all the islands involved there are three
A wiring can be in-segment:
This means that the slot assigned to the output port still belongs to the current segment. This is easy to detect
as the akka.stream.impl.IslandTracking class tracks the offset of the current segment. If the target input
slot is larger or equal than this offset, and the wiring is backwards, then the wiring is strictly local to the
island. The materializer will simply delegate to the akka.stream.impl.PhaseIsland to do the internal wiring.
Since we know the offset of the segment in the local space of this island, calculating the local slot for the
akka.stream.impl.PhaseIsland is simple. (This is fully documented with diagrams
A wiring can be cross-segment, in-island:
In this case, the target slot is in another, but already visited segment. The akka.stream.impl.IslandTracking
class needs to first find the segment in which the target slot is. Since each segment keeps a reference to its
akka.stream.impl.PhaseIsland instance that handles the internal wiring a simple reference equality check
will tell us if the target segment is in the same island or not. In this case it is, so all we need is to
compensate for any possible holes (punched by enclosed islands) to calculate the local slot for the island
and call the appropriate callback on the akka.stream.impl.PhaseIsland. (This is fully documented with diagrams
Finally a wiring can be cross-segment, cross-island:
This means, that the steps were similar as in the previous case until that point where we check the reference
equality of the current akka.stream.impl.PhaseIsland with that of the target segment (we have already found
the target segment). In this case, we need to calculate the local slot in the target island (similar to the
previous case) and try to wire the two islands together. Now, instead of delegating the wiring to the phases, we
ask the output akka.stream.impl.PhaseIsland to provide a Publisher and then we ask the target island to
take this Publisher.
Refer to akka.stream.impl.IslandTracking for all the nasty details of local slot resolution. It is also
recommended to try out a few examples with akka.stream.impl.PhasedFusingActorMaterializer.Debug turned on, it
will detail every step of the island tracking and slot resolution steps.
Useful utilities are:
* akka.stream.impl.PhasedFusingActorMaterializer.Debug: if this flag is turned on, the materializer will
log the steps it takes
* akka.stream.impl.TraversalBuilder.printTraversal: Prints the Traversal in a readable format
* akka.stream.impl.TraversalBuilder.printWiring: Prints the calculated port assignments. Useful for
debugging if everything is wired to the right thing.