Orchestration¶
Core Idea¶
Orchestration is coordinated execution under a directing control locus that selects, assigns, orders, and adapts heterogeneous components so their joint behavior realizes an outcome no component owns alone. [1]
The load-bearing question is where the process state lives. One logical place holds the answer to how far the execution has got, what each component returned, which deadlines are running, and what happens next — and that same place issues the next instruction. Components are not required to know they belong to anything larger; each receives an invocation, does its bounded work against its own contract, and returns. The plan, the accumulated results, the timers, the retry counters, and the exception paths all sit with the locus.
This is a narrower claim than the everyday word. Six commitments must hold together: two or more differentiated components, a directing locus carrying system-level state, an execution plan or policy, authority to assign and order, active management of the interfaces between components, and an integrated outcome that is nobody's individual output. [2]
The locus may be physically replicated — three engine replicas behind a leader election, a conductor with two assistants, a divisional staff cell — so long as one coherent control state resolves the execution. Logical singularity is the requirement; physical singularity is not.
Structural Signature¶
Held execution state → selective assignment → interface mediation → observed return → adapted next step. The cycle repeats until a terminal condition, and every arrow passes through the same locus. [3]
Recurring features:
- Component heterogeneity. Two or more participants expose different operations, contracts, resources, or effects. Were they interchangeable, plain dispatch would suffice and no directing model would be needed.
- A durable control state. Somewhere there is a readable record of where the execution has reached, and it outlives any single invocation.
- Assignment and ordering authority. The locus decides which component acts, supplies or routes its inputs, and fixes order, timing, or conditional activation — including the decision not to activate at all.
- Interface mediation. Components hold contracts with the locus rather than with each other, so the locus absorbs translation, formatting, unit conversion, and protocol mismatch.
- A closed return path. Outputs and observed conditions flow back into the control state, so later choices depend on what actually happened rather than on what was planned. [3]
- Exception and substitution authority. Handoffs, conflicts, timeouts, failures, retries, and replacements are the locus's business, not a concern bolted on afterwards.
- A non-owned outcome. The result has an identity at the level of the whole execution, and no participant can claim it.
What It Is Not¶
The word gets stretched in several directions this entry does not follow.
It does not mean being in charge. Setting objectives, owning budgets, approving plans, and reading reports can all be present while nobody holds the execution state. Deciding that a migration will happen this quarter is not orchestrating it. Deciding that the schema change runs before the backfill, handing the backfill its cutoff timestamp, and holding the pager until it returns, is. [4]
It does not mean scale, prestige, or sophistication. A forty-line script that calls three tools, inspects their exit codes, retries one of them twice, and branches on the result satisfies every commitment. A thousand-person program in which each department reacts to whatever the previous department published, with no place able to state the program's current position, satisfies none of them. Nor does the label follow the tooling: buying a product with "orchestrator" in its name does not create the role, and a coordinator with a whiteboard and a phone can hold it completely.
It is not the same as automation. The locus can be a person, a committee, a script, a control loop, or a regulatory circuit; what matters is that decisions about order and assignment resolve in one place, not that they resolve without human involvement. A fully automated system in which each stage fires the next on its own has automated the work while dispersing the direction. [1]
Finally, the locus is not required to understand its components. It needs a model of their interfaces, their observable effects, and their failure behavior, not of their internals — directing black boxes is the normal case rather than a degraded one. And the label carries no verdict. An orchestrated execution can be slow, wasteful, or wrong; calling it orchestration says only where the direction sits.
Broad Use¶
Distributed software. Workflow engines that keep a durable execution log across service calls; saga coordinators driving multi-service business transactions; build servers deciding which stages run against which artifacts; container control planes that repeatedly compare a declared desired state against observed reality and issue the difference as instructions; tool-calling agent loops that select the next tool from the accumulated transcript. [5]
Logistics and construction. A freight control tower that books capacity, rebooks it when a vessel slips, and reconciles customs, drayage, and warehouse windows. A general contractor who cannot pour before the rebar inspection passes and who resequences the trades when the inspector fails a section. A materials planning run that releases work orders against a bill of materials and revises them when a supplier confirms short.
Military operations. A fires cell deconflicting artillery, aviation, and ground movement in time and airspace, holding the execution matrix that says which asset is committed to which target in which window, and re-tasking when a target moves or a battery goes black.
Cellular and endocrine regulation. Checkpoint machinery gates cell-cycle progression on the observed completion of replication and spindle attachment; a hypothalamic-pituitary axis releases tropic signals to peripheral glands, reads their products back, and modulates the next release. Differentiated components, a state-carrying locus, and adaptive assignment are all present with nobody occupying the role of a conductor. [6]
Clinical care. A resuscitation team leader who performs no procedure but assigns each one, holds what has been done, and re-prioritizes on each new vital sign. A transplant coordinator sequencing donor recovery, transport, recipient preparation, and theatre availability across institutions that share no common system.
Performance and production. A conductor holding a global time against which every section's local reading is corrected; a first assistant director calling the shooting order against weather, light, and actor availability; a record producer sequencing tracking, overdub, and mix sessions across players who never meet.
Organizational programs. A program office running a platform migration wave by wave, holding the cutover checklist, gating each wave on the previous wave's error rate, and reassigning teams when a wave stalls.
Clarity¶
The confusion this prime dissolves is between having a plan and having a place where the plan is resolved at run time. A diagram of steps, a signed process document, and a shared understanding of the order of operations are design-time artifacts; none of them answers the question of where this particular execution has got to right now and who decides its next move. Orchestration names the runtime answer, and the two are routinely conflated: teams show an architecture diagram with arrows and conclude that the arrows are being enforced somewhere. [2]
It also disarms the argument about whether a system is centralized. That argument mixes three separate questions — where computation happens, where authority sits, and where process state is resolved — and deadlocks because participants are answering different ones. This prime isolates the third. Work spread over ten thousand machines in nine regions is still orchestrated if one control state decides what runs next; work confined to a single machine is not orchestrated if that machine merely reacts to whatever arrives.
A third confusion it clears concerns half-finished work. When a process is interrupted midway, the residue belongs to whoever holds the execution state. Where no holder exists, the residue belongs to nobody, and the interruption is discovered late, from its consequences.
Manages Complexity¶
Identifying the locus lets you stop tracking pairwise relationships between components. Under peer alignment, every participant must maintain some model of the others: whose message means what, whose vocabulary is authoritative, who has already acted, who still needs telling. That bookkeeping grows with the square of the participant count and is itself distributed, so no single copy of it is complete or checkable. With a directing locus, each component holds exactly one relationship — its contract with the locus — and the locus holds one model per component. The burden collapses from an unbounded web to a state machine plus n interfaces. [7]
Several questions consequently stop needing investigation. Whether every participant received a notification is not a question, because participants are called rather than notified. Whether two components agree about what a term means is not a question, because they never speak to each other. Whether a component understands the overall goal is not a question, because it is never asked to. Whether the sequence is being followed is not a question, because only one thing sequences it.
The debugging economy is the sharpest form of this. Locating a stalled execution becomes a single lookup against the control state rather than a reconstruction from the partial logs of every participant, correlated after the fact by timestamps that may not agree.
Abstract Reasoning¶
The prime licenses a short diagnostic that runs on any substrate. First, ask whether some single place can say how far the execution has got without polling the participants. Second, ask whether that place issues the next action or merely records what others did. Third, ask whether what a component returned changes what happens next, or whether the run would proceed identically regardless. Three affirmatives identify orchestration; a negative on the first identifies its opposite. [1]
The inferences that follow are what make the classification worth making. Given three affirmatives, a specific pathology set can be predicted before any investigation: the locus is both a stoppage point and a throughput ceiling, its model of the components will drift from their actual condition, it will accumulate decision logic nobody intended to put there, and partially completed runs will need explicit undo because no component will undo itself. You then go looking for the compensating actions, the failover, the timeout policy, and the staleness bound, and treat their absence as a finding.
Given a negative on the first question, expect the mirror-image set instead: nobody can state whether the process completed, partial completions stay invisible until their consequences surface, and adding a step means editing every participant that must react to it. The two sets barely overlap, which is why naming the mechanism tells you which failures to go looking for.
Knowledge Transfer¶
What carries is the role set and the failure taxonomy attached to it: a locus that can stop, saturate, go stale, and accrete; components that cannot recover a process they never saw; a control state whose durability decides whether an interruption is survivable. An engineer who has debugged a stalled workflow engine and a physician who has run a resuscitation are looking at the same object and can trade diagnoses without translation. [8]
What does not carry is the cost and physics of the control channel, and this is precisely where borrowed remedies fail. Issuing an instruction costs microseconds in one substrate, an hour and a fuel burn in another, and an undirected diffusion delay in a third. Components report status on request in some substrates and are readable only through indirect proxies in others, which sets how stale the locus's model may be allowed to get. Some loci can be interrogated about what they are waiting for and some cannot be asked anything at all.
Reversibility is the least portable property of the lot. A database write can be rolled back, a shipment recalled at cost, and an administered drug, a poured slab, or a transcribed gene cannot be undone, which decides whether compensation is a design choice or an impossibility. The working rule is to carry the diagnostic across substrates and re-derive the remedies locally.
Examples¶
Formal/abstract¶
Consider a booking that must reserve a flight, a hotel room, and a car from three independent providers, none of which will hold a lock while the others decide. No shared transaction can make the three commitments atomic, so atomicity has to be constructed. The orchestrated formulation introduces a coordinator holding a state machine over the states RESERVING_FLIGHT, RESERVING_HOTEL, RESERVING_CAR, CONFIRMED and their compensation counterparts, together with the identifiers returned so far. It calls the flight provider and records its reservation identifier, calls the hotel provider and records its identifier, calls the car provider — and on refusal at the third step walks its own record backwards, cancelling against the hotel identifier and then the flight identifier before terminating in a failed-but-clean state. [5]
Stated abstractly, an execution is a function from a plan and an accumulated history to a next action, and this prime is the claim that the function is evaluated at one site with the whole history in hand. The event-reactive formulation of the same booking evaluates no such function anywhere: each provider's adapter fires on what was published, applying a rule projected onto its own local view, and the global history exists only as the join of those views — a join nobody computes and no reader can obtain in a single query.
Mapped back: the coordinator is the directing locus; the three providers are heterogeneous components differing in protocol, latency, and cancellation semantics; the state machine plus identifier list is the durable control state; the refusal at step three is the observed return that changes what happens next; and the reverse walk is exception authority exercised as compensation, which the event-reactive variant cannot exercise because nobody holds the list of what to compensate. The completed booking is the non-owned outcome: no provider produced it, and none can cancel it. [5]
Applied/industry¶
A trauma resuscitation runs on the same structure with none of the vocabulary. The team leader stands at the foot of the bed and by protocol touches the patient as little as possible. Around the bed are people whose skills do not substitute for one another: an airway clinician, two placing access lines, a nurse drawing and labelling, a radiographer, a scribe, a runner to the blood bank. The leader holds the state — airway secured or not, two lines in or one, blood ordered and how many units, imaging done, what has been given and when — and assigns each next action to a named person aloud, so that assignment and acknowledgement both enter the record. When blood pressure drops after the initial survey, the leader does not restart the protocol from the top: the returned observation changes the next assignment, and imaging is deferred behind transfusion.
The failure modes are the predicted ones. A leader who picks up a scalpel has left the locus unattended and the room's state is no longer held anywhere. A leader receiving more reports than they can integrate becomes the throughput ceiling, and people start acting without assignment. A team member who acts without calling it out leaves the leader's model stale, which is why closed-loop callback is drilled rather than assumed. [9]
Mapped back: the leader is the directing locus, and deliberate non-participation in procedures is what keeps the control state held rather than interrupted. The clinicians are heterogeneous components whose contracts run to the leader rather than to each other, which is why they need no shared plan among themselves. Spoken assignment and callback are the closed return path made audible because no other durable channel is fast enough. Deferring imaging behind transfusion is adaptive re-ordering on observed return. The resuscitated patient is the outcome no participant owns.
Structural Tensions¶
T1 — The locus is a stoppage point as well as a control point. Every commitment that makes this prime legible also concentrates risk. Because one place resolves the execution, that place failing does not degrade the system gracefully; it halts it, and halts it mid-flight, with participants holding partial work they cannot resolve alone. Replication and failover shrink the outage but move the problem into the control state's consistency, since two loci disagreeing about the process position are worse than one that is simply down. The same concentration makes the locus a throughput ceiling long before any component saturates.
T2 — Visibility is bought with coupling. The reason to adopt this arrangement is that one query answers where the execution stands. The price is that every component must be reachable, addressable, and callable on the locus's terms, binding participants to a controller they did not choose and cannot version independently. Each new participant adds an integration at the centre rather than at the edge, and each change to a component's contract becomes a change to the controller. Answerability and independence trade against each other directly here, and no amount of engineering makes the trade disappear.
T3 — Ordering a step is not the same as being able to un-order it. When step four fails after steps one through three have committed, the locus must synthesize an undo the components never offered: cancellations, reversing entries, apology paths, clinical corrections. These compensations are business decisions dressed as error handling, they are the least exercised path in the system, and their correctness is hardest to establish precisely because they run only once something has already gone wrong. Where the substrate is irreversible, no compensation exists and the locus must instead refuse to begin.
T4 — The sequencer becomes the rule-keeper. A directing locus is introduced to say what happens next, not to decide what is true. But the conditions it branches on are substantive rules, and each new branch migrates one rule out of a component and into the centre. Over time the thing meant to hold order also holds eligibility, pricing, escalation, and exception policy, while components thin into effectors. The result is legible and centrally changeable, which is why it happens, and it becomes the hardest part to test or own because it now encodes everything the system knows.
T5 — Freshness at the centre versus freshness at the edge. The locus acts on a model of its components, and that model is always slightly out of date. Tightening it means polling, heartbeats, or mandatory callbacks: control traffic that consumes the very attention and bandwidth the arrangement was meant to conserve, and that scales with the participant count. Loosening it means assigning against a picture that has moved — work sent to a component already saturated, a step ordered against a precondition that has just lapsed. Every such system picks a staleness budget, usually implicitly, and inherits the misassignments it permits.
T6 — Adaptation versus replayability. The prime requires that the locus adapt, but discretion during the run is bought from predictability before it. A plan that fully determines its execution can be reviewed, tested, certified, and replayed; a locus permitted to re-sequence on what it observes produces runs the plan does not describe, and two runs of the same request may legitimately differ. Auditors, regulators, and incident reviewers then have to ask what the system did rather than what it does, which is answerable only if the control state recorded enough to reconstruct each decision.
Structural–Framed Character¶
Orchestration sits at the structural end of the structural–framed spectrum, graded structural with an aggregate of 0.00 and every criterion at zero. What travels is a control relation: two or more heterogeneous components exposing different operations, a single logical directing locus holding the execution-level objective and a model of those components, authority to select, route, order, and conditionally activate them, a return path so observed state informs later choices, and management of cross-component handoffs, conflicts, and failures — yielding an outcome no component owns alone.
The criterion doing the most work is human-practice-bound, and it reads zero because multicellular regulation instantiates the full role set with no actor in it: a directing locus with system-level state, differentiated components, and adaptive assignment, all present with nobody conducting. The same roles carry software workflows, logistics, military operations, and musical production.
Institutional origin reads zero for the same reason: no podium, command structure, or management office is required. Vocabulary travels at zero, the load-bearing terms being component, controller, sequence, and interface. Evaluative weight reads zero, since the prime is separated from its neighbours structurally rather than by merit: choreography coordinates through local rules with no directing controller, coordination can be emergent, scheduling orders work without assigning roles or managing interfaces. Import-vs-recognize is recognition.
The grade sets the test for applying the prime. A fixed list of steps with no runtime selection or adaptation is sequencing or a pipeline; independently controlled actors aligning through protocols or conventions are coordinating. Without a directing locus where joint execution is resolved, the label does not apply.
Substrate Independence¶
Orchestration is a highly substrate-independent prime — composite 4 / 5 on the substrate-independence scale. The invariant is a control topology: differentiated components, one logically singular locus holding the execution state, authority to assign and order, interfaces mediated so components contract with the locus rather than each other, and returns closing back into that state. The same topology carries workflow engines, logistics control, staff-directed military operations, hormonal regulation across a multicellular body, organizational programme management, and a recording session run from the desk. What keeps it a rung down is the commitment to a directing locus: an integrated outcome reached without one is a different prime rather than a weak instance, and the clean cases cluster where a designer installed the centre on purpose.
- Composite substrate independence — 4 / 5
- Domain breadth — 4 / 5
- Structural abstraction — 4 / 5
- Transfer evidence — 4 / 5
Relationships to Other Abstractions¶
Current abstraction Orchestration Prime
Foundational — no parent edges in the catalog.
Children (4) — more specific cases that build on this
-
Concert Band Domain-specific is a kind of Orchestration
a conductor and score select, assign, coordinate, and adapt heterogeneous instrumental resources into one performance; this is the proposed minimal parent.a conductor and score select, assign, coordinate, and adapt heterogeneous instrumental resources into one performance; this is the proposed minimal parent.
-
KLM protocol Domain-specific is a kind of Orchestration
The proposed strict upward parent is
prime:orchestration.prime:orchestration is the nearest broader Prime while the source-domain carrier and invariant supply the autonomous residual. This is a proposal-only workspace relationship: the accepted Prime supplies a genuinely instantiated structural prerequisite or superclass, while KLM protocol adds domain-specific constraints. The entry does not collapse into that parent because the domain-specific identity fixed by the photonic qubit encoding, single-photon and ancilla resources, linear optical elements, detectors and feedforward, nondeterministic gate, teleportation construction, success probability, error correction and resource scaling are explicit It also declines a nearby thematic catalog node: the neighbor does not literally subsume the constitutive identity of KLM protocol. This explicit assert-and-decline pattern keeps the proposed DAG narrow and prevents a merely thematic edge. The prospective workspace queue contains one strict upward edge toprime:orchestration. No live DAG mutation is authorized. -
Offstage instrument or choir part in classical music Domain-specific is a kind of Orchestration
The proposed strict upward parent is
prime:orchestration.The technique literally assigns and times a heterogeneous musical component as part of one execution while adding source location and acoustic path as orchestration variables. The edge is proposal-only and points to a frozen prior-baseline Prime. The entry does not collapse into the parent because the score-defined separation of a performing source from the main stage plus its acoustic and temporal coordination, not backstage ambience, incidental noise, surround playback, or a separate simultaneous concert A thematic neighbor is declined whenever it does not literally subsume that rule. The prospective workspace queue contains one strict upward edge toprime:orchestration. No live DAG mutation is authorized.
- Multi-Instrument Coordinated Campaign Domain-specific is a decomposition of Orchestration
A multi-instrument coordinated campaign is orchestration specialized to a directing strategist that adaptively assigns and sequences heterogeneous pressure instruments against a resistant target's defensive seams.Remove campaigns, strategy, contestation, adversaries, pressure instruments, defensive silos, escalation, defeat in detail, tobacco control, civil rights, corporate rivalry, and military examples. One logical control locus owns a cross-component execution objective and state, selects heterogeneous capabilities, routes their effects and observations, orders or conditionally activates them, and adapts the remaining execution after each result.
Neighborhood in Abstraction Space¶
Orchestration sits in a moderately populated region (48th percentile for distinctiveness): it has near-neighbors but no dense thicket of synonyms.
Family — Cue-Triggered Deferred Action (27 primes)
Nearest neighbors
- Interface Mismatch — 0.72
- Coordination — 0.71
- Boundary State Loss — 0.71
- Two-Store Architecture — 0.71
- Symmetric Response to Asymmetric State — 0.71
Computed from structural-signature embeddings · 2026-09-10
Not to Be Confused With¶
Orchestration must first be distinguished from Coordination, the broader relation it is one mechanism for achieving. Coordination names any arrangement in which independently controlled actors' separate actions combine into a coherent collective outcome, and it is silent about how that coherence arises: prices, conventions, protocols, shared intent, and mutual observation all produce it with nobody directing anything. Orchestration is the special case where coherence is produced by a place that decides. The practical consequence is asymmetric. Every orchestrated system is coordinated, but treating a merely coordinated system as orchestrated sends you hunting for a controller that does not exist, and treating an orchestrated one as merely coordinated hides the single participant whose failure stops everything.
The sharpest boundary runs against Self-Organization and, in software vocabulary, against the event-reactive pattern usually called choreography, whose canonical machinery is Publish–Subscribe. There, participants react to events they observe under local rules, with an anonymous broker routing but not deciding. The distinction is not about how many machines are involved or how code is deployed; it is about who holds the process state. An orchestrated execution has a reader that can be asked its position. A choreographed one has no such reader: the global position exists only as the join of every participant's local view, and reconstructing it means asking all of them and trusting that none has moved on since. Everything else follows from that single difference — who can compensate, who can time out, who can add a step, who is the bottleneck. A broker is not a locus; it moves messages and holds no opinion about how far anything has got.
Scheduling, Sequencing, and Pipeline each capture a part of this prime and stop short of it. Scheduling organizes work over time, deciding when and on what resource a set of tasks runs, without necessarily assigning differentiated roles, mediating between incompatible interfaces, or reacting to what a task returned. Sequencing fixes an order under precedence constraints, and a pipeline fixes a series of stages through which everything flows identically; both are properties of an arrangement rather than the activity of a directing site. A fixed order executed with no runtime selection, no conditional activation, and no exception authority is exactly what these three primes name. This prime begins where the order can change because of what just happened.
Composition makes a whole out of parts, and it is the structural cousin most often mistaken for orchestration because both yield something no part owns. Composition, though, is a statement about how parts are put together — the whole's structure is the composition — while orchestration is a statement about how parts are driven. Composed components can be entirely static: a compound expression, an assembled machine, an arranged image. Nothing in composition requires a site that decides which part acts next, and nothing in it addresses failure partway through assembly. A composed system that is never executed is still composed; an orchestration that is never executed is only a plan.
Mission Command and Delegation of Authority occupy the opposite pole on the authority axis while sharing the centralized-intent half. Mission command deliberately holds the why at the centre and pushes the how to the edge, on the argument that the edge has fresher information; delegation transfers a decision right to a new holder. In both, the recipient genuinely determines its own execution, and the centre neither knows nor fixes the steps taken. Orchestration keeps step-level determination at the locus. The two can coexist in one organization at different scales — theatre-level mission command over subordinate commanders each of whom orchestrates their own supporting arms — which is why the question must be asked about a specific level rather than about an institution.
Finally, Controllability is a property, not a role. It asserts that a system can be driven from any state to any other through the inputs available, and says nothing about whether a controller exists, what it holds, or how it decides. A system can be controllable with no orchestrator anywhere, and an orchestrated system can be built over components that are not controllable in this sense at all. Confusing the two produces a common error: concluding that because a system is steerable in principle, somebody must be steering it.
Solution Archetypes¶
No catalogued solution archetypes reference this prime yet.
References¶
[1] Peltz, Chris. "Web Services Orchestration and Choreography". Computer, 2003. Defines orchestration as an executable process that controls the flow from a single party's perspective, in explicit contrast with choreography, where no party holds the process and each acts on its own view of the exchange. registry ↩a ↩b ↩c
[2] Workflow Management Coalition. The Workflow Reference Model. Document TC00-1003, Issue 1.1, 1995. Separates build-time process definition from run-time enactment, and specifies an engine holding control data that identifies the state of each process and activity instance, navigating between activities and invoking both applications and human participants. registry ↩a ↩b
[3] Hohpe, Gregor, and Bobby Woolf. Enterprise Integration Patterns: Designing, Building, and Deploying Messaging Solutions. Addison-Wesley, 2003. The Process Manager pattern places a central unit that holds the state of the sequence, receives each reply back, and determines the next processing step from the accumulated intermediate results. registry ↩a ↩b
[4] Mintzberg, Henry. "Structure in 5's: A Synthesis of the Research on Organization Design". Management Science, 1980. Distinguishes direct supervision — one party issuing instructions and monitoring the work as it runs — from action planning and performance control, which set objectives and read results without touching the execution. registry ↩
[5] Garcia-Molina, Hector, and Kenneth Salem. "Sagas". Proceedings of the ACM SIGMOD International Conference on Management of Data, 1987. Introduces the coordinator that records completed sub-transactions and, on failure, walks its own record backwards issuing compensating actions. registry ↩a ↩b ↩c
[6] Hartwell, Leland H., and Ted A. Weinert. "Checkpoints: Controls That Ensure the Order of Cell Cycle Events". Science, 1989. Establishes checkpoint control that gates progression on observed completion of prior events rather than on elapsed time, with no actor occupying the directing role. registry ↩
[7] Gamma, Erich, Richard Helm, Ralph Johnson, and John Vlissides. Design Patterns: Elements of Reusable Object-Oriented Software. Addison-Wesley, 1994. The Mediator pattern replaces many-to-many interaction among peers with one relationship per participant once a mediator holds the interaction. registry ↩
[8] Skeen, Dale. "Nonblocking Commit Protocols". Proceedings of the ACM SIGMOD International Conference on Management of Data, 1981. Shows that a coordinator failing mid-protocol leaves participants blocked, holding partial work they cannot resolve among themselves, so the durability of the control state decides whether an interruption is survivable. registry ↩
[9] American College of Surgeons Committee on Trauma. Advanced Trauma Life Support: Student Course Manual. 10th ed. American College of Surgeons, 2018. Appendix E, on trauma team resource management, codifies a leader who assigns named roles and does not perform clinical procedures, and two-way communication in which the leader confirms back each item of information received. registry ↩