Release history

Changes to SupplyNetPy, newest first. The library was first published on PyPI on 2025-08-21 (version 0.1.0).

Each entry is tagged: [NEW] a new feature, [FIX] a bug fix, [CHANGE] a change in behavior, [BREAKING] a change that may need action from you, and [DOCS] documentation only.

0.5.0 (2026-10-17)

Simulation results are the same as in 0.4.0. This release adds probability-based supplier selection, and supplier-selection policies can now take settings.

[NEW] SelectProbabilistic, a supplier-selection policy that chooses the supplier of each replenishment order at random, with a probability for each supplier. The probabilities are given in the new site setting supplier_selection_param, keyed by supplier ID, and must sum to 1. When a supplier is disrupted, the others share its orders in their ratios (on_disruption="rescale", the default), or the order is blocked ("block"). The policy draws one random number per order from its own random number generator (seed or rng), so its choices change no other random numbers. create_sc_net checks that every link to the site has a probability. See the Supplier Selection guide.
[CHANGE] A site gives its supplier-selection policy a mode only when supplier_selection_mode is given, so each policy keeps its own default: "fixed" for the four existing policies, as before, and "dynamic" for SelectProbabilistic. Results are the same as in 0.4.0.
[BREAKING] A custom supplier-selection policy whose own __init__ takes mode without a default value now fails when the site is given no mode. Give mode a default, such as mode="fixed". Policies that do not override __init__ are not affected.
[CHANGE] A built in supplier-selection policy given settings it does not use in supplier_selection_param warns about them, as replenishment policies do for policy_param. A manufacturer given a supplier-selection policy, a mode or settings warns that they are not used, since it does not use its policy yet.
[CHANGE] get_sc_net_info shows the settings of each site's replenishment policy and supplier-selection policy, not only their names, for example (s: 100, S: 150, first_review_delay: 0, period: 0) or the probabilities of SelectProbabilistic. A distributor or retailer now reports its replenishment policy in get_info(), as a manufacturer already did.
[DOCS] The Supplier Selection guide has two new sections. One shows how to choose suppliers at random with SelectProbabilistic, and the other how to give settings to a supplier-selection policy you write yourself (self.params, _param_keys and check_links). New glossary entries for probability-based supplier selection and selection probability. The Known Issues page says that a manufacturer does not use its supplier-selection policy yet.
[DOCS] Level 3 headings in the guides and other pages are shown as plain headings. The grey box is kept for the classes and functions of the API reference.
[DOCS] The source code shown in the API reference keeps its line breaks and indentation again.

0.4.0 (2026-10-16)

Simulation results are the same as in 0.3.0. This release adds statistics for each link, and a User Guide with seven guides to common tasks.

[NEW] Every link now keeps statistics, in link.stats. They count the replenishment orders the buyer placed over the link, what arrived, the shortage and backorders at the supplier, how long orders waited at the supplier, the lead time, the costs, and the disruptions of the link. From these they work out the average waiting time, lead time and order cycle time, the fill rate (the share of units the supplier sent without the order having to wait), the availability of the link, and two spreads of order sizes, one per order and one per time unit, for studying the bullwhip effect. Read them with link.stats.get_statistics().
[NEW] get_link_wise_performance, format_link_wise_performance and print_link_wise_performance give the statistics of several links as data, as text or printed, like the node versions. Pass node.suppliers to see what a site ordered and from whom, or node.buyers to see who ordered from it.
[NEW] reset_stats(supplychainnet) and Network.reset_stats() start the statistics of every node, customer demand and link again in one call, for a warm-up.
[FIX] A raw-material supplier given a perishable inventory (inventory_type="perishable") now raises a clear error when it is created. SupplyNetPy does not model perishable raw materials yet. Before, the setting was accepted, and the run failed at the first mining with a message about a missing manufacturing date.
[CHANGE] SelectAvailable, SelectCheapest and SelectFastest (and the base class SupplierSelectionPolicy) now default to fixed mode, as SelectFirst and every site already did. A site always passes its own supplier_selection_mode, so networks give the same results as before. Only a policy created directly, without a mode, behaves differently.
[DOCS] A new User Guide with seven guides, on customer demand, replenishment policies, supplier selection, disruptions, perishable products, statistics and results (including link statistics and warm-ups), and running experiments. The longer sections of the API introduction moved into them, and the introduction is now a short overview of the modules.
[DOCS] New glossary entries for order cycle time, fill rate, bullwhip effect, standard deviation, fixed and dynamic mode, and, in a new Experiments section, replication, seed, confidence interval and common random numbers. Also a new example, examples/py/link_statistics.py.
[DOCS] The About page describes the newer features (customer demand options, supplier selection modes, perishable products, link statistics and reproducible experiments), links to the guides, and mentions the validation against anyLogistix. The home page lists the case studies, the User Guide, the known issues and the release history.
[DOCS] The planned work on the known issues page is split into next updates (probability-based supplier selection, orders in smaller parts, checking the whole network, and the new entry for perishable raw materials) and long-term goals.

0.3.0 (2026-10-16)

This release adds new node statistics, fixes how statistics are reset for a warm-up, and adds warnings for two mistakes in how a network is built. Results stay the same for networks that are not reset during a run and have no disrupted site with customers, such as the quickstart example. The three fixes below say which results change.

[FIX] Customers can no longer order from a disrupted site. Before, they were served from its inventory as usual, although a disrupted site takes no new replenishment orders. A customer who arrives while the site is disrupted now goes away, and the order is recorded in the new statistic disruption_lost_sales. It is kept apart from shortage, because a shortage is units the site did not have, and a disruption lost sale is an order the site could not take. Customers already waiting when the disruption starts are still served. Results change for networks in which a site with customers can be disrupted.
[FIX] stats.reset(), used for a warm-up, now starts the counting again without changing the present. Before, it also set backorder and inventory_level to zero. A site with orders waiting at the reset then ordered too little, and its backorders went below zero once those orders were served. The holding cost is now brought up to date before the reset, so afterwards it covers only the time since the reset, also when the run was advanced with env.run(). The record of inventory levels (inventory.instantaneous_levels) also starts again. Results after a warm-up change when orders are waiting at the reset, or when the run was advanced with env.run().
[FIX] A manufacturer's raw-material order now counts in its demand_placed and transportation_cost when it is placed, as for every other site. Before, it counted only once the supplier served it. Totals at the end of a run change only for raw-material orders still waiting then.
[NEW] Statistics about disruptions. disruptions and disrupted_time (including a disruption still going on when read), availability (the share of time the node was not disrupted), and blocked (replenishment orders a site could not place because every supplier it has, or the link to it, was disrupted).
[NEW] Statistics about inventory over time. avg_inventory_level and inventory_level_time (inventory level times time), worked out exactly and apart from the holding cost, and zero_inventory_time and zero_inventory_share (how long the inventory level was zero).
[NEW] Statistics about orders that end badly or wait. lost_sales (the customer units a site never delivered), dropped_qty (delivered units that did not fit in the inventory), and waited, waiting_time and avg_waiting_time, for customers and for orders from other sites.
[NEW] Each node's statistics keep a clock. start_time is when counting began, and elapsed_time is the time since then. A reset moves start_time, so averages after a warm-up cover only the measured period. Averages are worked out when they are read and are None when there is nothing to divide by.
[NEW] simulate_sc_net and Network.results add up lost_sales and disruption_lost_sales for the whole network.
[NEW] Node.has_suppliers() tells whether a site has any supplier at all. For a manufacturer it checks that every raw material has one.
[CHANGE] get_statistics() and the performance tables show the new statistics, so they have more rows.
[CHANGE] Two new warnings for mistakes in how a network is built. A site with no supplier at all that tries to order is warned about once, and is not counted as blocked. A buyer that can order more than its supplier can hold is warned about when the link between them is created, and again when such an order is placed. A supplier sends an order only when it has all of it, so such an order can never be filled, and the buyer would wait for it for ever. The warning names both sites.
[DOCS] New glossary entries for warm-up, disruption lost sales, lost sales, blocked order, waiting time, dropped units, average inventory level, availability and deadlock. The known issues page explains the deadlock above and lists two planned features, receiving such orders in smaller parts, and checking the whole network.

0.2.2 (2026-10-13)

Simulation results are the same as in 0.2.1. This release adds warnings for mistakes in how a model is set up, and makes the source package smaller.

[NEW] A replenishment policy now warns about keys in policy_param that it does not use, and asks whether another policy was meant. For example, 'T' is read only by the periodic policy, so passing it to the (s, S) policy now gives a warning instead of being silently ignored.
[CHANGE] Problems in how a model is set up are now shown with Python's warnings, so they appear even when logging is off. They are still written to the trace log. This covers the unused policy parameters above, a manufacturer without a supplier for every raw material, an unknown statistic name passed to update_stats, and these simulate_sc_net calls: a sim_time that is not later than the current time, logging=(start, stop) (deprecated, now a FutureWarning), and logging=(start, stop) together with log_window. Events during a run, such as a delivery that does not fit in the inventory, are still only logged.
[CHANGE] The source archive on PyPI now carries only the library, its tests, README, LICENSE and CHANGELOG, so it is about 115 KB instead of over 4 MB. The docs and examples stay in the repository. The wheel is unchanged.
[DOCS] The docs now say that a run stops just before its end time, so events due at exactly sim_time do not happen. This is how SimPy works. To include them, run to a slightly later time, such as sim_time + 1.

0.2.1 (2026-10-12)

Results change for networks with disruptions. Networks without disruptions, such as the quickstart example, give the same results as before.

[FIX] A site whose replenishment order was blocked by a disruption now orders as soon as its supplier, or the link to it, recovers. Before, a replenishment policy without a review period could stop ordering for good. If the site had no units left, nothing woke the policy again, even after the supplier recovered.
[FIX] Supplier-selection policies now skip a supplier whose site is disrupted, not only one whose link is disrupted. Before, a site with two suppliers could send its order to the disrupted one and lose it, while the other supplier was working.
[CHANGE] place_order now checks that some supplier can take the order before choosing one. If none can, no order is placed. New methods Link.can_take_order, Node.can_place_order, Node.wake_replenishment and Node.add_buyer_link, and a new list Node.buyers of the links to a node's buyers.
[FIX] The performance table (get_node_wise_performance, format_node_wise_performance, print_node_wise_performance) now shows every metric that any node has. Before, it showed only the metrics of the first node in the list, so a raw-material supplier's mining figures were left out when another node came first. Two nodes with the same name now each get a column; the second is told apart by its ID.
[FIX] disable_logging and set_log_file now close the log file they stop writing to. Before, the file stayed open until the program ended, which caused a ResourceWarning and, on Windows, kept the file locked.

0.2.0 (2026-10-11)

Most of these fixes come from issue #1, reported by @ashwingaikwadiby. Results change for networks with manufacturers, raw-material suppliers with finite inventory, perishable goods, customers who wait, or replenishment orders that have to wait at a supplier. Simple networks, such as the quickstart example, give the same results as before.

[FIX] A replenishment order that has to wait at its supplier is now a backorder there only while it waits, and is no longer counted twice in the supplier's inventory position. Before, the supplier ordered too much, kept the extra units for the rest of the run, and its backorders were never cleared.
[FIX] A manufacturer now keeps making production batches until every planned unit is made, orders raw material for every planned unit, and carries on when units leave its full inventory. Before, it could stop producing for good, with raw material left unused or with orders waiting forever.
[FIX] A raw-material supplier with finite inventory starts mining again as soon as its inventory drops. Before, once its inventory had been full it never mined again, and manufacturers buying from it waited forever.
[FIX] Perishable goods. An order now gets the batches that are really in inventory when it is filled, oldest first. Expired units are thrown away at once and are no longer shipped to an order that was waiting, and destroyed units are removed at once too. Before, an order that had to wait could lose units, and expired units could be shipped.
[FIX] Customers who wait (tolerance above 0) now leave as soon as their tolerance runs out, and their backorders are cleared when they leave. All waiting customers are woken when units arrive, two waiting customers can no longer take the same units, and a partial delivery of 0 units no longer crashes.
[FIX] A waiting customer's shortage is recorded once, when the order arrives. Before, it was recorded again at every check, so shortage figures were too high.
[FIX] A supplier whose site is disrupted is no longer counted as short for orders it never receives.
[FIX] Inventory.destroy() now wakes a raw-material supplier (to mine again) and a manufacturer (to make more). It still does not wake a replenishment policy.
[NEW] Inventory.get() takes a site_order argument, and Inventory.waiting_orders() and Inventory.waiting_site_orders() report the units asked for by orders still waiting at an inventory.
[DOCS] The (R, Q) policy (RQReplenishment) never used safety_stock, so passing it had no effect. The documentation no longer says that it does.
[DOCS] A new Glossary page defines one word for each idea, and the documentation, docstrings and code comments now use those words. The API reference now lists every result of simulate_sc_net, and no longer shows raw markup from the docstrings.
[DOCS] Several documentation pages described their examples wrongly, and are corrected. The complex network example now has the links shown in its picture, lead times that differ as intended, and a fixed seed.

0.1.12 (2026-05-31)

[FIX] Replenishment orders that were bigger than the free space on the shelf no longer throw away the extra units. Previously, a node with waiting customers could order enough to refill the shelf and also cover them, and any units that did not fit were silently lost even though they had been paid for. In some cases this also pushed the node's inventory position below zero.
[CHANGE] When arriving units are dropped because the shelf is full, a warning is now logged suggesting that the node's capacity may be too small.
[CHANGE] A customer order that is bigger than the node's whole shelf, with no partial delivery accepted, is now recorded as lost demand right away. Before, it triggered a replenishment order that could never be served.
[DOCS] Code comments and docstrings were rewritten in plainer language.

0.1.11 (2026-05-09)

[FIX] Random node and link disruptions set with failure_p no longer happen almost immediately at time 0. The failure chance is now applied once per time unit, as documented.

0.1.10 (2026-05-06)

[NEW] Inventory.destroy() removes stock (all of it, a fraction, or a custom amount) when a node is disrupted. Nodes accept a matching disruption_impact option. Destroyed units and their value appear in the node's statistics and count towards its cost.
[FIX] Expiry of perishable stock (Inventory.remove_expired) now works correctly.
[DOCS] Documentation wording was simplified throughout.

0.1.9 (2026-04-28)

[NEW] Network wrapper around the supply-chain dictionary, with named access to nodes and results. The dictionary-based way of working still works.
[NEW] NodeType, a list of the valid node types. Node types can be given either as text ("retailer") or as a NodeType member.
[NEW] log_window=(start, stop) option on simulate_sc_net to log only part of a run.
[BREAKING] Python 3.9 or newer is now required. networkx and matplotlib are now declared as dependencies.
[CHANGE] Logging is now opt-in. Creating nodes no longer opens (or overwrites) the trace file. simulate_sc_net(logging=True) turns it on for you.
[CHANGE] The shortage statistic is now called shortage. The old name, orders_shortage, still works as an alias.
[CHANGE] Expiry of perishable stock is now driven by the batches' expiry times instead of repeated checks, which makes large simulations faster.
[CHANGE] The remaining items from the code review (see REVIEW.md) were resolved, and the test suite was expanded.

0.1.8 (2026-04-24)

[NEW] set_seed() makes random disruptions repeatable.
[BREAKING] SupplyNetPy.Components now exposes a fixed, documented list of names. Modules that used to be reachable by accident (such as scm.simpy) are no longer available through it.
[FIX] A full code review fixed a set of correctness problems, including:
- Expired perishable stock was not always removed correctly, so holding cost could keep being charged on stock that no longer existed.
- A manufacturer could place duplicate orders for the same raw materials.
- A customer's "demand fulfilled" figure was inflated when an order arrived in several partial shipments.
- Cost totals could pick up unrelated values whose names happened to contain the word "cost".
- Disrupted links were still used for new shipments, and supplier-selection policies did not take link status into account.
- Adding stock to a full shelf could leave the inventory records out of step.
[CHANGE] Misspelled node parameters (for example shelf_live) are now reported as an error instead of being silently ignored.
[CHANGE] A much larger test suite was added.

0.1.7 (2026-03-13)

[NEW] record_inv_levels option on Inventory to switch recording of inventory levels over time on or off (off by default).
[FIX] Adding stock to an inventory that is already full is now handled safely.

0.1.6 (2025-11-19)

[FIX] When adding stock would go over capacity, only the amount that fits is now added, instead of failing.

0.1.5 (2025-11-08)

[NEW] name is now included in node statistics.
[CHANGE] The minimum required Python version was relaxed.

0.1.2 to 0.1.4 (2025-09-15)

[FIX] The wheel build target was added so the package builds correctly.
[CHANGE] Package metadata (classifiers and author details) was corrected.

0.1.1 (2025-09-02)

[CHANGE] Parameters that expect a "callable" (lead times, arrival intervals, and so on) now also accept a plain number such as 2 or 1.5.

0.1.0 (2025-08-21)

First public release on PyPI. It included:

Supply-chain building blocks: suppliers, manufacturers, inventory nodes (distributors and retailers), demand nodes, and links.
Replenishment policies: (s, S), (R, Q), and periodic review (with or without safety stock).
Supplier-selection policies: first, available, cheapest, and fastest.
Perishable inventory, disruptions of nodes and links, and backorders with partial fulfilment.
Node-level and network-level results (cost, revenue, profit, shortage), plus create_sc_net and simulate_sc_net.
A documentation site, worked examples, and a unit-test suite.

Before the first release

Development began in May 2024; the repository was moved from its earlier home in July 2024. Internal versions 0.0.11 to 0.0.69 were used during development and were never published on PyPI.