DD Pipeline Two-Phase Architecture: Client/Loader Pattern

Due diligence investigations require orchestrating dozens of data sources with wildly different response times, reliability characteristics, and rate limits. A naive sequential approach collapses under real-world conditions. This post details the two-phase Client/Loader architecture that powers Prismatic's DD pipeline, enabling concurrent data acquisition with structured entity processing.

The Problem with Monolithic Pipelines

Traditional DD platforms fetch data sequentially: query a business registry, wait for a response, query a court database, wait again, then attempt to stitch results together. This approach has three fatal flaws. First, total latency equals the sum of all source latencies, meaning a 15-source investigation can take minutes. Second, a single source failure blocks the entire pipeline. Third, there is no natural point for incremental result delivery to the user.

Phase 1: The Client Layer

The Client layer is responsible for concurrent data acquisition from external sources. Each source is wrapped in an adapter module implementing the Prismatic.DD.Source behaviour:

defmodule Prismatic.DD.Source do
@callback fetch(entity :: map(), opts :: keyword()) ::
{:ok, %{data: map(), confidence: float(), source: atom()}}
| {:error, term()}


@callback source_group() :: :registry | :court | :financial | :sanctions | :media
@callback rate_limit() :: {requests :: pos_integer(), window_ms :: pos_integer()}
@callback priority() :: 1..10
end

Source groups are a critical design element. Rather than treating all sources equally, we partition them into logical groups: :registry for business registries, :court for litigation databases, :financial for financial data providers, :sanctions for sanctions lists, and :media for news and media monitoring. This grouping drives the scheduling strategy.

Concurrent Fetching with Backpressure

The Client layer uses Task.async_stream/3 with controlled concurrency per source group:

defmodule Prismatic.DD.Client do
@moduledoc """
Concurrent data acquisition layer for DD pipeline.
Manages source groups with independent concurrency limits.
"""


@group_concurrency %{
registry: 5,
court: 3,
financial: 4,
sanctions: 10,
media: 8
}


@spec fetch_all(map(), [module()], keyword()) :: [source_result()]
def fetch_all(entity, sources, opts \\ []) do
timeout = Keyword.get(opts, :timeout, 30_000)


sources
|> Enum.group_by(& &1.source_group())
|> Enum.flat_map(fn {group, group_sources} ->
max_concurrency = Map.get(@group_concurrency, group, 3)


group_sources
|> Task.async_stream(
fn source -> execute_with_telemetry(source, entity, opts) end,
max_concurrency: max_concurrency,
timeout: timeout,
on_timeout: :kill_task
)
|> Enum.map(fn
{:ok, result} -> result
{:exit, :timeout} -> {:error, :timeout}
end)
end)
end


defp execute_with_telemetry(source, entity, opts) do
start_time = System.monotonic_time()


:telemetry.execute(
[:prismatic, :dd, :source, :start],
%{system_time: System.system_time()},
%{source: source, entity_id: entity.id}
)


result = source.fetch(entity, opts)


:telemetry.execute(
[:prismatic, :dd, :source, :stop],
%{duration: System.monotonic_time() - start_time},
%{source: source, entity_id: entity.id, result: elem(result, 0)}
)


result
end
end

Each source group operates with its own concurrency limit. Sanctions lists are highly parallelizable (stateless HTTP lookups), so they get 10 concurrent slots. Court databases are more fragile, limited to 3. This prevents any single group from starving others.

Rate Limiting

Rate limiting is implemented per-source using a token bucket stored in ETS:

defmodule Prismatic.DD.RateLimiter do
@moduledoc """
Token bucket rate limiter backed by ETS for sub-millisecond access.
"""


use GenServer
require Logger


def check_rate(source) do
{max_requests, window_ms} = source.rate_limit()
bucket_key = {source, current_window(window_ms)}


case :ets.update_counter(@table, bucket_key, {2, 1}, {bucket_key, 0}) do
count when count <= max_requests -> :ok
_over_limit -> {:error, :rate_limited}
end
end


defp current_window(window_ms) do
div(System.system_time(:millisecond), window_ms)
end
end

Phase 2: The Loader Layer

Once raw data arrives from the Client layer, the Loader layer takes over. Its responsibilities are entity resolution, data normalization, confidence scoring, and persistence. The Loader operates as a GenStage pipeline with three stages: normalizer, resolver, and persister.

Entity Resolution

The most complex part of the Loader is entity resolution. The same company might appear as "Acme s.r.o.", "ACME, s.r.o.", or "Acme spol. s r.o." across different registries. The resolver uses a combination of exact ICO matching, fuzzy name matching with Jaro-Winkler distance, and address normalization:

defmodule Prismatic.DD.Loader.Resolver do
@moduledoc """
Entity resolution across heterogeneous data sources.
"""


@jaro_threshold 0.92


@spec resolve(raw_entity :: map(), existing :: [map()]) :: {:match, map()} | {:new, map()}
def resolve(raw_entity, existing_entities) do
cond do
match = find_by_ico(raw_entity, existing_entities) ->
{:match, merge_entity(match, raw_entity)}


match = find_by_fuzzy_name(raw_entity, existing_entities) ->
{:match, merge_entity(match, raw_entity)}


true ->
{:new, normalize_entity(raw_entity)}
end
end


defp find_by_ico(%{ico: ico}, entities) when is_binary(ico) do
Enum.find(entities, fn e -> e.ico == ico end)
end


defp find_by_ico(_entity, _entities), do: nil


defp find_by_fuzzy_name(%{name: name}, entities) when is_binary(name) do
normalized = normalize_name(name)


Enum.find(entities, fn e ->
String.jaro_distance(normalized, normalize_name(e.name)) >= @jaro_threshold
end)
end


defp find_by_fuzzy_name(_entity, _entities), do: nil


defp normalize_name(name) do
name
|> String.downcase()
|> String.replace(~r/\s(s\.r\.o\.|spol\.\ss\sr\.o\.|a\.s\.|s\.p\.)\s/i, "")
|> String.replace(~r/[,\.\-]+/, " ")
|> String.trim()
end
end

PubSub Streaming

As the Loader processes entities, results are broadcast via Phoenix PubSub on the "dd:pipeline" topic. This enables real-time LiveView updates without polling:

defmodule Prismatic.DD.Loader.Broadcaster do
@moduledoc """
PubSub broadcasting for DD pipeline progress and results.
"""


@topic "dd:pipeline"


@spec broadcast_progress(case_id :: binary(), event :: atom(), payload :: map()) :: :ok
def broadcast_progress(case_id, event, payload) do
Phoenix.PubSub.broadcast(
Prismatic.PubSub,
"#{@topic}:#{case_id}",
{event, Map.put(payload, :timestamp, DateTime.utc_now())}
)
end


@spec broadcast_entity_resolved(case_id :: binary(), entity :: map()) :: :ok
def broadcast_entity_resolved(case_id, entity) do
broadcast_progress(case_id, :entity_resolved, %{
entity_id: entity.id,
name: entity.name,
source_count: length(entity.sources),
confidence: entity.confidence
})
end
end

LiveView components subscribe to case-specific topics and render results incrementally as they arrive. The user sees entities appearing in real-time rather than waiting for the entire pipeline to complete.

The Scheduler

Orchestrating the two phases is the DD Scheduler, a GenServer that manages pipeline lifecycle:

defmodule Prismatic.DD.Scheduler do
@moduledoc """
Pipeline lifecycle management for DD investigations.
"""


use GenServer
require Logger


def start_investigation(case_id, entities, opts \\ []) do
GenServer.call(__MODULE__, {:start, case_id, entities, opts})
end


@impl true
def handle_call({:start, case_id, entities, opts}, _from, state) do
sources = select_sources(entities, opts)


task =
Task.Supervisor.async_nolink(Prismatic.DD.TaskSupervisor, fn ->
results = Prismatic.DD.Client.fetch_all(entities, sources, opts)


Prismatic.DD.Loader.process(case_id, results, opts)
end)


new_state = Map.put(state, case_id, %{task: task, started_at: DateTime.utc_now()})
{:reply, {:ok, case_id}, new_state}
end
end

The Scheduler uses Task.Supervisor.async_nolink/2 to isolate pipeline failures. If a pipeline crashes, it does not bring down the Scheduler, and other running investigations continue unaffected.

Performance Characteristics

In production, the two-phase architecture achieves median investigation times of 4.2 seconds for a standard 15-source Czech company DD, compared to 45+ seconds with sequential fetching. The 90th percentile sits at 8.1 seconds, bounded primarily by the slowest external source.

The key insight is that the Client/Loader separation creates a natural boundary for error handling. Client-layer failures (network timeouts, rate limits) are transient and retryable. Loader-layer failures (entity resolution conflicts, schema violations) are logical and require investigation. Mixing these concerns in a monolithic pipeline makes error recovery vastly more complex.

Conclusion

The two-phase architecture transforms DD from a batch process into a streaming operation. Users see results within seconds rather than minutes. The system gracefully degrades when sources are unavailable. And the clear separation between data acquisition and entity processing makes the codebase maintainable as we add new sources. This pattern has proven robust across thousands of investigations and continues to scale with the platform.