article addressing performance concerns with the current schema

docs/articles/tradeoffs.md

@@ -0,0 +1,167 @@
+---
+title: Performance Trade-offs in the TROLIE Schema Design
+parent: Articles
+---
+
+{: .nb }
+> At the time of the this article, only the Forecast portion of the
+> TROLIE 1.0 specification is considered stable. However, the design choices made
+> for Forecasts are anticipated to be used in RT and Seasonal as well.
+
+# Performance Trade-offs in the TROLIE Schema Design
+
+## Introduction
+
+Initial feedback from the broader electric power industry community is that the
+JSON format that TROLIE currently specifies is going to be inefficient. Some are
+concerned that if they adopted that schema, they wouldn't be able to meet their
+performance objectives, given the volume of ratings submissions they will be
+handling.
+
+The goals of this article are:
+
+* Discussing the relative priority of certain quality attributes in the TROLIE
+  project
+
+* Addressing the rationale for the current schema specification
+
+* Mitigating performance concerns of implementors
+
+## Quality Attribute Priorities
+
+TROLIE primarily exists to foster interoperability between systems that need to
+exchange ratings and limits in order to comply with FERC Order 881. Accordingly,
+the project prioritizes qualities that enhance interoperability and security
+over other qualities attributes related to performance efficiency.
+
+Attributes like clarity, compatibility/flexibility, debuggability, auditability,
+and [robustness](https://en.wikipedia.org/wiki/Robustness_principle) promote
+interoperability, yet these attributes are often in tension with performance
+efficiency. Consideration of these attributes led to the choice of boring,
+long-established technologies--JSON, HTTP, OpenAPI--to specify TROLIE. Security
+concerns like auditability and non-repudiation have also informed the schema
+design.
+
+## Schema Design Rationale
+
+The entire schema for [Forecast
+Proposals](https://trolie.energy/spec#tag/Forecasting/operation/patchRatingForecastProposal)
+is certain worth scrutinizing, but for our immediate purposes, let's dilate on
+just the `ratings` object of the message.
+
+```jsonc
+{
+  "proposal-header": { /* details elided for clarity */ },
+  "ratings": [
+    {
+      "resource-id": "8badf00d",
+      "periods": [
+        {
+          "period-start": "2025-11-01T01:00:00-05:00",
+          "period-end": "2025-11-01T02:00:00-05:00",
+          "continuous-operating-limit": {
+            "mva": 160
+          },
+          "emergency-operating-limits": [
+            {
+              "duration-name": "lte",
+              "limit": {
+                "mva": 165
+              }
+            },
+            {
+              "duration-name": "ste",
+              "limit": {
+                "mva": 170
+              }
+            },
+            {
+              "duration-name": "dal",
+              "limit": {
+                "mva": 170
+              }
+            }
+          ]
+        } // other periods elided for clarity
+      ]
+    } // other resource forecasts elided for clarity
+  ]
+}
+```
+
+There are a few considerations to call out here:
+
+* _Does not assume that rating applies to a single hour._ While that
+ is a fine assumption to make for most if not all current Order 881
+ implementations, the current TROLIE approach allows for flexibility in how the
+ interval for a given rating is applied. Since this approach is used in Forecast
+ and RT Ratings Proposals as well as the Limits Snapshots, implementors benefit
+ from uniformity in API. (Exchange of DLRs may become necessary
+ [soon](https://www.ferc.gov/news-events/news/sunshine-notice-june-2024-commission-meeting).)
+
+* _Does not assume emergency rating durations._ Not all entities utilize the
+  same emergency durations, and while a given TO to RC exchange might be able to
+  make this assumption, an interop standard has to be compatible with all RCs.
+
+* _Does not assume units for ratings._ Not all entities utilize the same units
+  for specifying limits, of course. The TROLIE schema allows for different units
+  to be used with different power system resources in the same forecast.
+
+## Mitigating Performance Concerns
+
+Still, the performance concerns are real. This article is not an attempt to 
+discredit them. Let's address the one's we've heard.
+
+### Concern 1: Transferred Bytes
+
+JSON is a *terrible* choice from a
+performance perspective, but it is also ubiquitous and will be so for decades to
+come. It also has a decent schema language that OpenAPI builds upon.
+
+Certainly, though, a more compact way to represent the data we need to exchange,
+even in bloated JSON, is possible. However, JSON compresses *extremely* well,
+and much of the data in the proposed format is repeated, like the timestamps
+required for each period. Gzip is perhaps more ubiquitous than JSON, and the
+widely implemented Brotli is incredibly efficient at compressing JSON.
+
+Moreover, the semantics of PATCH (used to submit proposals) allow a Ratings
+Provider to break up their submissions into multiple requests if they chose to
+employ that as a tactic to ensure robustness of the exchange, i.e., to support
+smaller payloads that are faster to retry.
+
+### Concern 2: Processing Efficiency
+
+There are some obvious performance considerations when implementing any exchange
+that has to support 881, because
+
+> 240 hours × multiple emergency durations × a large footprint = a lot of ratings to process.
+
+#### Don't Read It All Into Memory
+
+Implementors are strongly encourages to use a streaming or event-based
+("SAX-style" for you XML aficionados) parser when reading submissions. This
+advice would likely apply in a general way, even given a ratings exchange
+specification that made different trade-off decisions. TROLIE specifies a
+`proposal-header` that afford the server the opportunity to make some processing
+decisions early, without waiting to receive the last byte.
+
+#### Ignore What You Don't Need
+
+Also, implementors are encouraged to *ignore what you don't need*. For example,
+it is definitely the case that most submissions will have 240 periods. While the
+TROLIE schema requires `period-start` and `period-end` timestamps for each one,
+if you are in a position to dictate conventions for the exchange, you could
+ignore these and require a submission always include 240 sequential hourly
+period forecasts; doing so would not violate the TROLIE schema, and you would
+still have an unambiguous record of the intended use of the rating if it comes
+to non-repudiation or debugging.
+
+## Interop is a Conversation
+
+The Maintainers believe the current schema for Forecast Proposals and Forecast
+Limits Snapshot strikes a good balance between performance and fostering
+interoperability, but nothing speaks louder than data. To that end we hope to
+update this post soon with actual code and performance data, but if you need to
+make a go/no-go decision about TROLIE today, and this is the issue holding you
+back, please don't let it. The schema is not written in stone and better ideas
+are always welcome.