[Canonical page]() [Full documentation index]() # Leveling Resources This guide covers how to remove resource overallocation from a plan: what [`Project::level`]() changes, how it picks the task that gives way, how to limit it to part of the plan, and how to read, preview and undo its result. ## How leveling works A resource is overallocated when the tasks booked on it need more than its capacity at the same time. [`solve`]() only reports this, as `overallocations` windows on the result. [`level`]() fixes it by moving work later. It never changes assignments, resources or capacities. It changes two task fields: - `leveling_delay`, which delays the whole task. - `pauses`, which splits a task: it stops working for a while and picks up again later. Leveling looks at the plan in **buckets**: day buckets when `config.duration_unit` is `Day`, hour buckets for `Hour` and `Minute`. The **load** of a resource in a bucket is the sum of `value` over the assignments that book working time in that bucket. A task counts with its full `value` as soon as it works any time in the bucket. When the load is higher than the resource's `capacity` (default `config.default_resource_capacity`, which is 1), the bucket is overallocated. A run of overallocated buckets in a row is a **window**. `level` works through the plan from left to right, in rounds: 1. Find the earliest overallocated bucket. 2. Collect the tasks that contribute to it, the **contenders**, and sort them by the selection order. 3. Move the weakest contender just past the bucket, with a delay or a pause, until the load fits. 4. Reschedule the project, so links and constraints push the successors along. 5. Repeat until no overallocated bucket with a movable contender is left. Some tasks never move. Leveling leaves these alone and calls them the **fixed set**: - tasks with `auto_scheduling: false` - completed tasks (`progress == 1`, when `config.use_progress` is on) - tasks with an `Mso` or `Mfo` constraint (when `config.apply_constraints` is on) - tasks with `priority == 1000` - summaries (their dates follow their children) - tasks outside `LevelingScope.tasks`, when you set it If a proposed move doesn't stick, because a constraint or another hard rule pulls the task back, leveling reverts it and freezes the task for the rest of the run. Hard requirements always win. Whatever leveling can't resolve stays in `overallocations`, and each remaining window gets a `reason`. A few rules hold for every run: - `level` takes `&mut self` and commits. It solves the current state first, writes the delays and pauses into the project and accepts the leveled dates as the new baseline. To show leveling as a proposal, run it on a clone. See [Previewing and clearing](). - Leveling is forward-only. A project with `config.schedule_from_end: true` is refused with [`Error::Input`](). - The output is deterministic. Every tie ends in input order, so the same data and options always give the same result. - There's no round limit. A run ends when nothing is left to resolve. On large projects it can take a while, so run it on a worker thread in a server. ## Leveling a project Call [`level`]() with [`LevelingOptions::default()`]() after the project is loaded. The examples on this page use the default calendar (Monday to Friday, 08:00-12:00 and 13:00-17:00) and start on Monday, August 10, 2026. ```rust use plen_core::{ AssignmentData, DateTime, LevelingOptions, Project, ProjectData, ResourceData, SolveOptions, TaskData, }; fn aug(day: u32) -> Result> { Ok(DateTime::from_ymd_hm(2026, 8, day, 8, 0).ok_or("invalid date")?) } fn task(id: &str, start: DateTime, duration: f64) -> TaskData { let mut task = TaskData::new(id); task.start_date = Some(start); task.duration = Some(duration); task } fn booking(task: &str, resource: &str) -> AssignmentData { AssignmentData { id: None, task: task.into(), resource: resource.into(), value: 1.0, offset: None, duration: None, } } fn main() -> Result<(), Box> { // Two 2-day tasks, both booked full time on dev1 from Monday. let data = ProjectData { tasks: vec![task("A", aug(10)?, 2.0), task("B", aug(10)?, 2.0)], resources: vec![ResourceData { id: "dev1".into(), capacity: None, calendar_id: None, }], assignments: vec![booking("A", "dev1"), booking("B", "dev1")], ..Default::default() }; let mut project = Project::default(); let solved = project.solve(Some(data), &SolveOptions::default())?; for over in solved.overallocations.iter().flatten() { for window in &over.windows { // dev1: 2026-08-10 08:00 .. 2026-08-11 17:00, peak 2 of 1 println!( "{}: {} .. {}, peak {} of {}", over.resource, window.from, window.to, window.peak, window.capacity ); } } let leveled = project.level(&LevelingOptions::default())?; for update in &leveled.schedule.updates { if let (Some(start), Some(end)) = (update.start_date, update.end_date) { // B: 2026-08-12 08:00 .. 2026-08-14 08:00, delay Some(2.0) println!( "{}: {start} .. {end}, delay {:?}", update.id, update.leveling_delay ); } } Ok(()) } ``` [`solve`]() reports one window on `dev1` over Monday and Tuesday, with a peak load of 2. `level` keeps `A` and moves `B` to Wednesday. Both tasks are equal, so the input-order tie-break decides: the later task moves. Leveling diffs its `updates` against the state the call started from, so you can apply them to the model you already hold. Next to the new dates, each entry carries the task's new `leveling_delay` and, when it changed, its complete `pauses` list. An update entry maps field-for-field onto a [`TaskUpdate`](). ## Delays and pauses `leveling_delay` and `pauses` are **elapsed** time, in the task's duration unit: with day units, 1 is 24 hours of calendar time, weekends and nights included. `duration` and `lag` stay working time. - A delay shifts the task's start from its driven position (the start that links and constraints give it). Draw it as the stretch `[start - delay, start)` before the bar. - A `Pause { at, length, by }` starts `at` after the task start and lasts `length`. The leveler sets `by` to `Some(PauseSource::Leveling)` on every pause it writes. Draw pauses as gaps in the bar. They need no calendar. Which one leveling picks depends on the contender. A task that starts inside the overallocated bucket has no work before it, so it gets a whole-task delay. A task that's already running in that bucket can be paused: it keeps the work it has done and resumes after the contention. Pausing a running task leaves no idle hole behind it. Delaying it would. Set `allow_splits: false` to get whole-task delays only. ```rust use plen_core::{ AssignmentData, DateTime, LevelingOptions, LinkData, LinkType, Project, ProjectData, ResourceData, SolveOptions, TaskData, }; fn aug(day: u32) -> Result> { Ok(DateTime::from_ymd_hm(2026, 8, day, 8, 0).ok_or("invalid date")?) } fn task(id: &str, start: DateTime, duration: f64) -> TaskData { let mut task = TaskData::new(id); task.start_date = Some(start); task.duration = Some(duration); task } fn booking(task: &str, resource: &str) -> AssignmentData { AssignmentData { id: None, task: task.into(), resource: resource.into(), value: 1.0, offset: None, duration: None, } } fn resource(id: &str) -> ResourceData { ResourceData { id: id.into(), capacity: None, calendar_id: None, } } fn main() -> Result<(), Box> { // A runs Monday and Tuesday. B joins on Tuesday and feeds C, so it has less slack. let data = ProjectData { tasks: vec![ task("A", aug(10)?, 2.0), task("B", aug(11)?, 2.0), task("C", aug(13)?, 5.0), ], links: vec![LinkData { id: "bc".into(), source: "B".into(), target: "C".into(), link_type: LinkType::Fs, lag: 0.0, lag_unit: None, }], resources: vec![resource("dev1"), resource("dev2")], assignments: vec![ booking("A", "dev1"), booking("B", "dev1"), booking("C", "dev2"), ], ..Default::default() }; let mut project = Project::default(); project.solve(Some(data), &SolveOptions::default())?; // Splits allowed (the default): A is paused, B and C keep their dates. let mut split = project.clone(); let result = split.level(&LevelingOptions::default())?; for update in &result.schedule.updates { // A: end Some("2026-08-14 08:00"), pauses [(1.0, 2.0)] let pauses: Vec<(f64, f64)> = update .pauses .iter() .flatten() .map(|p| (p.at, p.length)) .collect(); println!( "{}: end {:?}, pauses {pauses:?}", update.id, update.end_date.map(|d| d.to_string()) ); } // Whole-task delays only: B moves past A, and C moves with it. let mut whole = project.clone(); let options = LevelingOptions { allow_splits: false, ..Default::default() }; let result = whole.level(&options)?; for update in &result.schedule.updates { if let (Some(start), Some(end)) = (update.start_date, update.end_date) { // B: 2026-08-12 08:00 .. 2026-08-14 08:00 // C: 2026-08-14 08:00 .. 2026-08-21 08:00 println!("{}: {start} .. {end}", update.id); } } Ok(()) } ``` With splits, `A` keeps its start and works Monday, pauses for two days (`at: 1.0`, `length: 2.0`), and finishes its second day on Thursday. `B` works in the pause, so the chain into `C` doesn't move. Without splits, `A` can't be paused and stays in place, so `B` is delayed whole and the link pushes `C` along. The update for `A` has no `start_date`, because its start didn't change. Pauses only come from leveling: there are no user-defined pauses. A pause takes up no load, and a task without pauses never gets a `pauses` field in its update. The field rules are on the [`Pause` reference](). ## Leveling options [`LevelingOptions`]() implements `Default`. Override fields with struct update syntax: | Field | Default | What it changes | |---|---|---| | `scope` | whole project | Which resources, tasks and dates leveling acts on. | | `basis` | derived from `config.duration_unit` | Bucket size, [`Basis::Day`]() or `Basis::Hour`. | | `order` | [`LevelingOrder::Standard`]() | Which contender moves first. See [Selection order](). | | `within_slack_only` | `false` | Never push a task past the project end of the plan leveling started from. | | `clear_first` | `true` | Clear the existing delays and pauses on in-scope tasks before leveling. | | `allow_splits` | `true` | Allow `pauses`. With `false`, every move is a whole-task delay. | | `step` | [`LevelingStep::Minimal`]() | How far one move goes. | The full list is on the [`LevelingOptions` reference](). ### Leveling part of a plan [`LevelingScope`]() limits a run in three ways, and you can combine them: - `tasks`: only these tasks may move. Everything else acts like a fixed task. - `resources`: only contention on these resources triggers a move. - `from` / `to`: only buckets that start in `[from, to)` trigger a move. The range limits where leveling looks for contention, not where tasks land: a moved task may end up past `to`. ```rust use plen_core::{ AssignmentData, DateTime, LevelingOptions, LevelingScope, Project, ProjectData, ResourceData, SolveOptions, TaskData, }; fn aug(day: u32, hour: u32) -> Result> { Ok(DateTime::from_ymd_hm(2026, 8, day, hour, 0).ok_or("invalid date")?) } fn task(id: &str, start: DateTime, duration: f64) -> TaskData { let mut task = TaskData::new(id); task.start_date = Some(start); task.duration = Some(duration); task } fn booking(task: &str, resource: &str) -> AssignmentData { AssignmentData { id: None, task: task.into(), resource: resource.into(), value: 1.0, offset: None, duration: None, } } fn main() -> Result<(), Box> { let data = ProjectData { tasks: vec![task("A", aug(10, 8)?, 2.0), task("B", aug(10, 8)?, 2.0)], resources: vec![ResourceData { id: "dev1".into(), capacity: None, calendar_id: None, }], assignments: vec![booking("A", "dev1"), booking("B", "dev1")], ..Default::default() }; let mut project = Project::default(); project.solve(Some(data), &SolveOptions::default())?; // Only A may move, so A gives way even though B is later in input order. let only_a = LevelingOptions { scope: Some(LevelingScope { tasks: Some(vec!["A".into()]), ..Default::default() }), ..Default::default() }; let result = project.clone().level(&only_a)?; for update in &result.schedule.updates { if let (Some(start), Some(end)) = (update.start_date, update.end_date) { // A: 2026-08-12 08:00 .. 2026-08-14 08:00 println!("{}: {start} .. {end}", update.id); } } // Only contention from Wednesday on. The Monday-Tuesday window stays. let from_wednesday = LevelingOptions { scope: Some(LevelingScope { from: Some(aug(12, 0)?), ..Default::default() }), ..Default::default() }; let result = project.clone().level(&from_wednesday)?; for over in result.schedule.overallocations.iter().flatten() { for window in &over.windows { // dev1: 2026-08-10 08:00 .. 2026-08-11 17:00, Some("out-of-scope") println!( "{}: {} .. {}, {:?}", over.resource, window.from, window.to, window.reason ); } } Ok(()) } ``` The second run leaves the Monday window alone and reports it with the reason `out-of-scope`. The result still says `converged: true`, because the run did everything it was asked to do. The reason tells you why the window is still there. An unknown id in `scope.tasks` or `scope.resources` fails the call with [`Error::Input`](), at the path `scope.tasks` or `scope.resources`. ### Staying within slack With `within_slack_only: true`, leveling never pushes a task further than its total slack. The slack is measured against the project end of the plan the run started from, so leveling can't make the project finish later. If a contender can't clear the bucket within its slack, leveling skips it and tries the next one. If no contender fits, the window stays with the reason `slack-exhausted`. The slack leveling uses is internal. It ignores the existing delays and uses that fixed project end, so it can differ from `analytics.slack` in a [`solve`]() result. ### Fewer rounds with the skip step By default a move clears one bucket. Then leveling reschedules and re-selects at the next bucket, so the burden of a long window can switch between contenders bucket by bucket. `step: LevelingStep::Skip` moves the task past the whole window in one move. That takes fewer rounds and reschedules less often. Use it for large projects where [`level`]() is slow. ```rust use plen_core::{ AssignmentData, DateTime, LevelingOptions, LevelingStep, Project, ProjectData, ResourceData, SolveOptions, TaskData, }; fn aug(day: u32) -> Result> { Ok(DateTime::from_ymd_hm(2026, 8, day, 8, 0).ok_or("invalid date")?) } fn task(id: &str, start: DateTime, duration: f64) -> TaskData { let mut task = TaskData::new(id); task.start_date = Some(start); task.duration = Some(duration); task } fn booking(task: &str) -> AssignmentData { AssignmentData { id: None, task: task.into(), resource: "dev1".into(), value: 1.0, offset: None, duration: None, } } fn main() -> Result<(), Box> { let data = ProjectData { tasks: vec![task("A", aug(10)?, 2.0), task("B", aug(10)?, 2.0)], resources: vec![ResourceData { id: "dev1".into(), capacity: None, calendar_id: None, }], assignments: vec![booking("A"), booking("B")], ..Default::default() }; let mut project = Project::default(); project.solve(Some(data), &SolveOptions::default())?; let minimal = project.clone().level(&LevelingOptions::default())?; let skip = project.clone().level(&LevelingOptions { step: LevelingStep::Skip, ..Default::default() })?; // Same dates for B (Wednesday to Friday), in 2 rounds and in 1. println!( "minimal: {} rounds, skip: {} rounds", minimal.rounds, skip.rounds ); assert_eq!(minimal.schedule.updates, skip.schedule.updates); Ok(()) } ``` Here both steps give the same plan. That's not guaranteed: buckets the skip step jumps over never get a new selection, so on real plans the two can level differently. If a jump hits a constraint or doesn't fit the slack, leveling falls back to the one-bucket move for that task. ## Selection order When several contenders share an overallocated bucket, the order decides which one stays and which moves first. Every order ends in the same tie-break: the task later in input order moves. | [`LevelingOrder`]() | Who moves first | |---|---| | `Standard` (default) | Checked in order: a successor before its predecessor, more total slack, later start, lower `priority`, no dated constraint. | | `Priority` | Lower `priority` first, then the `Standard` checks. | | [`Id`]() | The task later in input order. | All three orders also apply the running rule. A task already working before the bucket stays, and the task that starts in the bucket moves, unless the running task can be paused instead. This keeps whole-task delays from leaving gaps in work that was already scheduled. Under `Priority`, a higher priority beats this rule: the lower-priority running task is paused, or delayed whole when splits are off. `priority` is a `u16` on [`TaskData`](), 500 by default. At 1000 the task never moves. The dependency check exists because delaying a predecessor would push its successor along and resolve nothing. In this plan `B` comes first in the input, but an SS link from `A` makes `A` its predecessor, so `B` moves: ```rust use plen_core::{ AssignmentData, DateTime, LevelingOptions, LinkData, LinkType, Project, ProjectData, ResourceData, SolveOptions, TaskData, }; fn aug(day: u32) -> Result> { Ok(DateTime::from_ymd_hm(2026, 8, day, 8, 0).ok_or("invalid date")?) } fn task(id: &str, start: DateTime, duration: f64) -> TaskData { let mut task = TaskData::new(id); task.start_date = Some(start); task.duration = Some(duration); task } fn booking(task: &str) -> AssignmentData { AssignmentData { id: None, task: task.into(), resource: "dev1".into(), value: 1.0, offset: None, duration: None, } } fn main() -> Result<(), Box> { let data = ProjectData { tasks: vec![task("B", aug(10)?, 2.0), task("A", aug(10)?, 2.0)], links: vec![LinkData { id: "ab".into(), source: "A".into(), target: "B".into(), link_type: LinkType::Ss, lag: 0.0, lag_unit: None, }], resources: vec![ResourceData { id: "dev1".into(), capacity: None, calendar_id: None, }], assignments: vec![booking("A"), booking("B")], ..Default::default() }; let mut project = Project::default(); project.solve(Some(data), &SolveOptions::default())?; let result = project.level(&LevelingOptions::default())?; for update in &result.schedule.updates { if let (Some(start), Some(end)) = (update.start_date, update.end_date) { // B: 2026-08-12 08:00 .. 2026-08-14 08:00 println!("{}: {start} .. {end}", update.id); } } Ok(()) } ``` To protect a task from leveling, set its `priority` to 1000 or its `auto_scheduling` to `false`. With `B.priority = 1000` in the first example, `A` moves to Wednesday instead. ## Reading the result [`level`]() returns a [`LevelingResult`](): - [`schedule`](): an `Arc`, the same shape `solve` returns. Its `updates` carry the new dates, delays and pauses. `baseline_id` names the new baseline. - `rounds`: the number of rounds that proposed a move. Rounds that only skipped an unresolvable bucket don't count. `schedule.converged` is the scheduler's convergence on the final plan, not a leveling flag. Leveling itself always runs until nothing is left to resolve. `schedule.overallocations` lists the windows that remain. Every window from `level` has a `reason`. When a window has several immovable contenders, the first reason in this table wins: | `reason` | Why the window remains | |---|---| | `pinned-constraint` | A contender was frozen because its move didn't stick, or it has an `Mso` / `Mfo` constraint. | | `fixed-task` | A contender has `auto_scheduling: false`, or is a summary. | | `completed` | A contender has `progress == 1` under `config.use_progress`. | | `priority-1000` | A contender has `priority == 1000`. | | `slack-exhausted` | Every candidate was capped by `within_slack_only`. | | `out-of-scope` | The resource, the dates or every contender is outside `scope`. | A reason is data, not an error. It tells the user what to change to resolve the window: unpin a task, loosen a constraint, widen the scope. ```rust use plen_core::{ AssignmentData, ConstraintType, DateTime, LevelingOptions, Project, ProjectData, ResourceData, SolveOptions, TaskData, }; fn aug(day: u32) -> Result> { Ok(DateTime::from_ymd_hm(2026, 8, day, 8, 0).ok_or("invalid date")?) } fn task(id: &str, start: DateTime, duration: f64) -> TaskData { let mut task = TaskData::new(id); task.start_date = Some(start); task.duration = Some(duration); task } fn booking(task: &str) -> AssignmentData { AssignmentData { id: None, task: task.into(), resource: "dev1".into(), value: 1.0, offset: None, duration: None, } } fn main() -> Result<(), Box> { // A is fixed. B is the only task that could move, but it must start by Monday. let mut a = task("A", aug(10)?, 2.0); a.auto_scheduling = false; let mut b = task("B", aug(10)?, 2.0); b.constraint_type = Some(ConstraintType::Snlt); b.constraint_date = Some(aug(10)?); let mut data = ProjectData { tasks: vec![a, b], resources: vec![ResourceData { id: "dev1".into(), capacity: None, calendar_id: None, }], assignments: vec![booking("A"), booking("B")], ..Default::default() }; data.config.apply_constraints = true; let mut project = Project::default(); project.solve(Some(data), &SolveOptions::default())?; let result = project.level(&LevelingOptions::default())?; // The move was reverted, so nothing changed. assert!(result.schedule.updates.is_empty()); for over in result.schedule.overallocations.iter().flatten() { for window in &over.windows { // dev1: tasks [A, B], reason Some("pinned-constraint") let tasks: Vec = window.tasks.iter().map(|id| id.to_string()).collect(); println!( "{}: tasks {tasks:?}, reason {:?}", over.resource, window.reason ); } } Ok(()) } ``` Leveling proposed a delay for `B`, the replan pulled `B` back to Monday, and leveling reverted the move. Nothing in the project changed. The window stays and reports `pinned-constraint`. With both tasks fixed and no constraint, the same window would report `fixed-task`. The other window fields (`from`, `to`, `capacity`, `peak`, `overload`, `tasks`) are on the [`OverallocationWindow` reference](). For the load of each bucket, see [Resource load](). `level` fails with [`Error::Input`]() for an unknown scope id or a backward-scheduled project, and with [`ProjectError::NeedsResync`]() after a rejected [`apply`](). Both leave the project as it was. See [Error handling](). ## Previewing and clearing Because [`level`]() commits, run it on a clone to show leveling as a proposal. [`Project::clone`]() is cheap: the data is shared until one copy is edited. ```rust use plen_core::{LevelingOptions, Project}; fn leveling_proposal(project: &Project) -> Result<(), Box> { let mut draft = project.clone(); let proposal = draft.level(&LevelingOptions::default())?; for update in &proposal.schedule.updates { println!( "{}: delay {:?}, pauses {:?}", update.id, update.leveling_delay, update.pauses ); } // Accept: keep `draft` in place of `project`. Reject: drop it. Ok(()) } ``` The step-by-step version is in [Preview leveling](). To score leveling as part of an alternative change set, use [`evaluate`]() with `EvaluateOptions.level`. See [Evaluating alternatives](). There are two ways to remove leveling: - To undo it, set `leveling_delay` to 0 and `pauses` to an empty list on the task, then [`solve`](). The original contention comes back. - To level again from scratch, call `level` again. `clear_first` is on by default, so the run first clears the delays and pauses of the in-scope tasks and levels from the clean plan. With `clear_first: false` the existing delays and pauses stay, and leveling adds to them. Use that to keep a deliberate gap the user set by hand. ```rust use plen_core::{ AssignmentData, DateTime, LevelingOptions, Patch, Project, ProjectData, ResourceData, SchedulingChanges, SolveOptions, TaskData, TaskUpdate, }; fn aug(day: u32) -> Result> { Ok(DateTime::from_ymd_hm(2026, 8, day, 8, 0).ok_or("invalid date")?) } fn task(id: &str, start: DateTime, duration: f64) -> TaskData { let mut task = TaskData::new(id); task.start_date = Some(start); task.duration = Some(duration); task } fn booking(task: &str) -> AssignmentData { AssignmentData { id: None, task: task.into(), resource: "dev1".into(), value: 1.0, offset: None, duration: None, } } fn main() -> Result<(), Box> { let data = ProjectData { tasks: vec![task("A", aug(10)?, 2.0), task("B", aug(10)?, 2.0)], resources: vec![ResourceData { id: "dev1".into(), capacity: None, calendar_id: None, }], assignments: vec![booking("A"), booking("B")], ..Default::default() }; let mut project = Project::default(); project.solve(Some(data), &SolveOptions::default())?; project.level(&LevelingOptions::default())?; // B moves to Wednesday // Undo leveling on B. let mut clear = TaskUpdate::new("B"); clear.leveling_delay = Patch::Value(0.0); clear.pauses = Patch::Value(Vec::new()); let mut changes = SchedulingChanges::default(); changes.tasks.update.push(clear); project.apply(&changes)?; let replanned = project.solve(None, &SolveOptions::default())?; for update in &replanned.updates { if let (Some(start), Some(end)) = (update.start_date, update.end_date) { // B: 2026-08-10 08:00 .. 2026-08-12 08:00 println!("{}: {start} .. {end}", update.id); } } // dev1 is overallocated on Monday and Tuesday again. assert!(replanned .overallocations .iter() .flatten() .any(|o| !o.windows.is_empty())); Ok(()) } ``` The `leveling_delay` update moves `B`'s stored start back by the same amount, so `B` returns to Monday and the window over Monday and Tuesday comes back. To clear every task, put one such update per task with a delay or pauses into the same change set. The full recipe is in [Clear leveling]().