Ostrzyciel · Oct 27, 2026 · Oct 27, 2026 · Oct 27, 2026
diff --git a/docs/performance/index.md b/docs/performance/index.md
@@ -1,7 +1,7 @@
 # Performance benchmarks
 
-[Jelly-JVM]({{ jvm_link() }}) 4.0.0 was benchmarked against the serialization formats built into Apache Jena and Eclipse RDF4J:
+[Jelly-JVM]({{ jvm_link() }}) 4.0.0 was benchmarked against the serialization formats built into Apache Jena and Eclipse RDF4J, and [pyjelly]({{ python_link() }}) 0.9.0 against the formats built into RDFLib:
 
-- **[Jelly-RDF](rdf.md)** – RDF triples and quads, compared with N-Triples / N-Quads, Turtle / TriG, RDF/XML, JSON-LD, RDF/JSON, Jena's binary formats, and RDF4J Binary.
-- **[Jelly-SPARQL](sparql.md)** – SPARQL query results, compared with the XML, JSON, and TSV results formats, Jena's binary formats, and RDF4J Binary.
-- **[Jelly-RDF 1.1 vs 1.2](rdf-1-1-vs-1-2.md)** – Jelly-RDF 1.2 (the default in Jelly-JVM 4.0.0) compared with Jelly-RDF 1.1.
+- **[Jelly-RDF](rdf.md)** – RDF triples and quads, compared with the formats in Apache Jena, Eclipse RDF4J, and RDFLib.
+- **[Jelly-SPARQL](sparql.md)** – SPARQL query results, compared with the formats in Apache Jena, Eclipse RDF4J, and RDFLib.
+- **[Jelly-RDF 1.1 vs 1.2](rdf-1-1-vs-1-2.md)** – Jelly-RDF 1.2 (the default in Jelly-JVM 4.0.0 and pyjelly 0.9.0) compared with Jelly-RDF 1.1.
diff --git a/docs/performance/rdf-1-1-vs-1-2.md b/docs/performance/rdf-1-1-vs-1-2.md
@@ -1,10 +1,17 @@
 # Jelly-RDF 1.1 vs 1.2
 
-Jelly-RDF 1.2 features a new column layout for stream frames, which improves throughput and reduces serialized size (see the [changes from Jelly-RDF 1.1](../specification/serialization.md#changes-from-jelly-rdf-11)). The first version of Jelly-JVM to support it is 4.0.0, which also includes other performance improvements in the implementation itself that made Jelly-RDF 1.1 faster as well. This page compares Jelly-RDF 1.2 with Jelly-RDF 1.1, and also shows the performance improvements coming from the implementation.
+Jelly-RDF 1.2 features a new column layout for stream frames, which improves throughput and reduces serialized size (see the [changes from Jelly-RDF 1.1](../specification/serialization.md#changes-from-jelly-rdf-11)). The first versions to support it are Jelly-JVM 4.0.0 and pyjelly 0.9.0. Both also include other performance improvements in the implementation itself that made Jelly-RDF 1.1 faster as well. This page compares Jelly-RDF 1.2 with Jelly-RDF 1.1, and also shows the performance improvements coming from the implementations.
 
 The charts are interactive – hover over them for details.
 
-## Benchmark setup
+**Jump to:**
+
+- **[Jelly-JVM](#jelly-jvm)** – Apache Jena and Eclipse RDF4J
+- **[pyjelly](#pyjelly)** – RDFLib and the generic API
+
+## Jelly-JVM
+
+### Benchmark setup
 
 The same datasets, hardware, software, and method were used as in the [Jelly-RDF benchmarks](rdf.md#benchmark-setup), with the "big" preset. Three versions were compared:
 
@@ -14,9 +21,9 @@ The same datasets, hardware, software, and method were used as in the [Jelly-RDF
 
 Each of the three versions was run 6 times for every dataset. The result for a dataset is the median of the 6 runs.
 
-## Results
+### Results
 
-### Throughput
+#### Throughput
 
 --8<-- "snippets/performance/e-versions.html"
 
@@ -32,7 +39,7 @@ Throughput on each of the 13 datasets (median of 6 runs). Log scale.
 
 --8<-- "snippets/performance/k-speed-variants.data.html"
 
-### Serialized size
+#### Serialized size
 
 --8<-- "snippets/performance/i-size-summary.html"
 
@@ -52,7 +59,7 @@ Serialized size on each of the 13 datasets, uncompressed and with zstd. Sorted b
 
 --8<-- "snippets/performance/j-size-variants.data.html"
 
-### Throughput and serialized size
+#### Throughput and serialized size
 
 --8<-- "snippets/performance/l-speed-size-variants.html"
 
@@ -61,8 +68,64 @@ Size (uncompressed) against throughput on each of the 13 datasets. Each line con
 
 --8<-- "snippets/performance/l-speed-size-variants.data.html"
 
+## pyjelly
+
+### Benchmark setup
+
+The same datasets, hardware, software, and method were used as in the [pyjelly Jelly-RDF benchmarks](rdf.md#pyjelly), with pyjelly's streaming API and the "big" preset. Throughput was measured with RDFLib terms and with the terms of pyjelly's [generic API]({{ python_link('generic-sink') }}), which does not use RDFLib. Three versions were compared:
+
+- **0.8 (Jelly-RDF 1.0)** – pyjelly 0.8.2, Jelly-RDF 1.0, frame size 250 (its defaults)
+- **0.9 (Jelly-RDF 1.1)** – pyjelly 0.9.0 writing Jelly-RDF 1.1, frame size 1024
+- **0.9 (Jelly-RDF 1.2)** – pyjelly 0.9.0, Jelly-RDF 1.2, frame size 1024
+
+Each of the three versions was run 6 times for every dataset. The result for a dataset is the median of the 6 runs. Serialized sizes are from the RDFLib integration (the generic API writes the same bytes in 0.9.0, and within 3% of them in 0.8.2).
+
+### Results
+
+#### Throughput
+
+--8<-- "snippets/performance/py-e-versions.html"
+
+Step-by-step throughput improvements: the code changes in pyjelly 0.9.0 (with Jelly-RDF 1.1), and the Jelly-RDF 1.2 format. Teal: geometric mean over 13 datasets, gray: each dataset.
+{ .jc-caption }
+
+--8<-- "snippets/performance/py-e-versions.data.html"
+
+--8<-- "snippets/performance/py-k-speed-variants.html"
+
+Throughput on each of the 13 datasets (median of 6 runs). Log scale.
+{ .jc-caption }
+
+--8<-- "snippets/performance/py-k-speed-variants.data.html"
+
+#### Serialized size
+
+--8<-- "snippets/performance/py-i-size-summary.html"
+
+Serialized size, geometric mean over 13 datasets.
+{ .jc-caption }
+
+--8<-- "snippets/performance/py-i-size-summary.data.html"
+
+--8<-- "snippets/performance/py-j-size-variants.html"
+
+Serialized size on each of the 13 datasets, uncompressed and with zstd. Sorted by how much smaller Jelly-RDF 1.2 is. Log scale.
+{ .jc-caption }
+
+--8<-- "snippets/performance/py-j-size-variants.data.html"
+
+#### Throughput and serialized size
+
+--8<-- "snippets/performance/py-l-speed-size-variants.html"
+
+Size (uncompressed) against throughput on each of the 13 datasets. Each line connects a dataset's results for 0.8 (Jelly-RDF 1.0) and 0.9 (Jelly-RDF 1.2). Up and to the left is better. Log scales.
+{ .jc-caption }
+
+--8<-- "snippets/performance/py-l-speed-size-variants.data.html"
+
 ## See also
 
 - [Jelly-RDF benchmarks](rdf.md)
 - [Jelly-RDF 1.2 specification](../specification/serialization.md)
-- [Benchmark code](https://github.com/Jelly-RDF/jelly-jvm/tree/main/jmh)
+- [Jelly-JVM benchmark code](https://github.com/Jelly-RDF/jelly-jvm/tree/main/jmh)
+- [pyjelly benchmark code](https://github.com/Jelly-RDF/pyjelly/tree/main/benchmarks)
diff --git a/docs/performance/rdf.md b/docs/performance/rdf.md
@@ -1,12 +1,19 @@
 # Jelly-RDF benchmarks
 
-The following results were obtained by benchmarking [Jelly-JVM]({{ jvm_link() }}) 4.0.0 against the RDF serialization formats built into Apache Jena and Eclipse RDF4J. For SPARQL query results, see [Jelly-SPARQL benchmarks](sparql.md).
+The following results were obtained by benchmarking [Jelly-JVM]({{ jvm_link() }}) 4.0.0 against the RDF serialization formats built into Apache Jena and Eclipse RDF4J, and [pyjelly]({{ python_link() }}) 0.9.0 against the formats built into RDFLib. For SPARQL query results, see [Jelly-SPARQL benchmarks](sparql.md).
 
 The benchmarks were performed on flat streams of RDF triples or quads.
 
 The charts are interactive – hover over them for details.
 
-## Benchmark setup
+**Jump to:**
+
+- **[Jelly-JVM](#jelly-jvm)** – Apache Jena and Eclipse RDF4J
+- **[pyjelly](#pyjelly)** – RDFLib
+
+## Jelly-JVM
+
+### Benchmark setup
 
 All benchmarks were performed on the first 100,000 statements of each dataset in [RiverBench](https://w3id.org/riverbench/v/dev) (development version, October 2026), except `yago-annotated-facts`, which uses RDF-star. Out of the 13 used datasets, 3 include RDF quads (`assist-iot-weather-graphs`, `citypulse-traffic-graphs`, `nanopubs`).
 
@@ -20,7 +27,7 @@ Hardware: NVIDIA DGX Spark (GB10 Grace Blackwell Superchip with 10 Cortex-X925 a
 
 Software: Linux kernel 6.17, GraalVM CE 25.4 (Java 25.0.4), Apache Jena 6.2.0, Eclipse RDF4J 6.1.0, [Jelly-JVM]({{ jvm_link() }}) 4.0.0.
 
-### Tested methods
+#### Tested methods
 
 - W3C N-Triples / N-Quads (Apache Jena, Eclipse RDF4J)
 - W3C Turtle / TriG (Apache Jena `TURTLE_BLOCKS` and `TRIG_BLOCKS`, Eclipse RDF4J without pretty-printing)
@@ -40,9 +47,9 @@ Software: Linux kernel 6.17, GraalVM CE 25.4 (Java 25.0.4), Apache Jena 6.2.0, E
 
     See the [dedicated Jelly-RDF 1.1 vs 1.2 comparison](rdf-1-1-vs-1-2.md) for details on the performance improvements.
 
-## Results
+### Results
 
-### Throughput
+#### Throughput
 
 --8<-- "snippets/performance/a-overview-rdf.html"
 
@@ -51,14 +58,14 @@ Throughput, geometric mean over 13 datasets (10 for RDF/JSON and RDF/XML, which
 
 <div class="jelly-chart-notes" markdown>
 
-- \* buffering writer: drops duplicate statements (write throughput still counts every input statement)
+- \* buffering writer: deduplicates statements (write throughput still counts every input statement)
 - ° RDF4J's writer is given an OutputStream and writes it one character at a time. With a BufferedWriter it could be 1.28–1.56× faster
 
 </div>
 
 --8<-- "snippets/performance/a-overview-rdf.data.html"
 
-### Throughput and serialized size
+#### Throughput and serialized size
 
 --8<-- "snippets/performance/b-speed-size-rdf-triples.html"
 
@@ -67,14 +74,14 @@ Serialized size (uncompressed) against throughput, geometric mean over 10 datase
 
 <div class="jelly-chart-notes" markdown>
 
-- \* buffering writer: drops duplicate statements (write throughput still counts every input statement)
+- \* buffering writer: deduplicates statements (write throughput still counts every input statement)
 - ° RDF4J's writer is given an OutputStream and writes it one character at a time. With a BufferedWriter it could be 1.28–1.56× faster
 
 </div>
 
 --8<-- "snippets/performance/b-speed-size-rdf-triples.data.html"
 
-### Throughput per dataset
+#### Throughput per dataset
 
 --8<-- "snippets/performance/c-heatmap-rdf.html"
 
@@ -83,7 +90,7 @@ Throughput on each dataset, in statements/s. The last 3 datasets contain quads.
 
 <div class="jelly-chart-notes" markdown>
 
-- \* buffering writer: drops duplicate statements (write throughput still counts every input statement)
+- \* buffering writer: deduplicates statements (write throughput still counts every input statement)
 - § the 3 sweeps differ by more than 10%
 - ° RDF4J's writer is given an OutputStream and writes it one character at a time. With a BufferedWriter it could be 1.28–1.56× faster
 
@@ -98,15 +105,15 @@ Throughput on each dataset (dots) and the geometric mean over 10 datasets (bar).
 
 <div class="jelly-chart-notes" markdown>
 
-- \* buffering writer: drops duplicate statements (write throughput still counts every input statement)
+- \* buffering writer: deduplicates statements (write throughput still counts every input statement)
 - § the 3 sweeps differ by more than 10%
 - ° RDF4J's writer is given an OutputStream and writes it one character at a time. With a BufferedWriter it could be 1.28–1.56× faster
 
 </div>
 
 --8<-- "snippets/performance/d-spread-rdf-triples.data.html"
 
-### Serialized size
+#### Serialized size
 
 --8<-- "snippets/performance/f-size-rdf-triples.html"
 
@@ -115,16 +122,125 @@ Serialized size in bytes/statement, uncompressed and with zstd, geometric mean o
 
 <div class="jelly-chart-notes" markdown>
 
-- \* buffering writer: drops duplicate statements (write throughput still counts every input statement)
+- \* buffering writer: deduplicates statements (write throughput still counts every input statement)
 
 </div>
 
 --8<-- "snippets/performance/f-size-rdf-triples.data.html"
 
+## pyjelly
+
+### Benchmark setup
+
+The same 13 datasets (first 100,000 statements of each) and the same hardware were used as for [Jelly-JVM](#benchmark-setup).
+
+- Each method/dataset combination was run in 3 separate Python processes (sweeps), each with 1 × 1 s warmup and 5 × 1 s measurement. The result for a dataset is the median of the 3 sweeps.
+- Serialization: the data was preloaded into memory as a list of RDFLib terms and serialized to a null output stream. RDFLib can write Turtle, RDF/XML, and JSON-LD only from a whole `Graph` or `Dataset`, so for these formats the graph was built beforehand, and only `serialize` was measured.
+- Deserialization: the data was parsed from bytes in memory. Only the parsed statements were counted in the throughput numbers.
+- Writers were given the prefixes of the dataset.
+- Averages are geometric means over the datasets.
+
+Hardware: the same NVIDIA DGX Spark, pinned to 5 Cortex-X925 cores. The benchmarks are single-threaded.
+
+Software: Linux kernel 6.17, CPython 3.15.1, RDFLib 7.6.0, protobuf 7.36.2 (upb backend), [pyjelly]({{ python_link() }}) 0.9.0 (compiled with mypyc).
+
+#### Tested methods
+
+- W3C N-Triples / N-Quads (RDFLib's serializer, run line by line over the list of statements)
+- W3C Turtle / TriG (RDFLib)
+- W3C RDF/XML (RDFLib) – triples only. RDFLib can't parse its own RDF/XML for 4 of the datasets, so there are results for 6 datasets.
+- W3C JSON-LD (RDFLib)
+- Jelly (pyjelly 0.9.0, Jelly-RDF 1.2, "big" preset), with pyjelly's streaming API (`flat_stream_to_file` and `parse_jelly_flat`)
+    ```protobuf
+    max_name_table_size = 4000;
+    max_prefix_table_size = 150;
+    max_datatype_table_size = 32;
+    ```
+- Jelly (Graph API) – the same, through RDFLib's plugin API (`Graph.serialize(format="jelly")` and `Graph.parse(format="jelly")`)
+
+RDFLib changes some literals when parsing them (e.g., `"73"^^xsd:float` is parsed as `"73.0"`), so N-Triples, Turtle, RDF/XML, and JSON-LD don't round-trip exactly on 7 of the 13 datasets. These results are still included. Jelly round-trips exactly on all datasets.
+
+### Results
+
+#### Throughput
+
+--8<-- "snippets/performance/py-a-overview-rdf.html"
+
+Throughput, geometric mean over 13 datasets (6 for RDF/XML). Log scale.
+{ .jc-caption }
+
+<div class="jelly-chart-notes" markdown>
+
+- \* writes from an RDFLib `Graph`, which deduplicates statements (throughput counts only the statements written)
+
+</div>
+
+--8<-- "snippets/performance/py-a-overview-rdf.data.html"
+
+#### Throughput and serialized size
+
+--8<-- "snippets/performance/py-b-speed-size-rdf-triples.html"
+
+Serialized size (uncompressed) against throughput, geometric mean over 10 datasets (6 for RDF/XML). Log scale. Top left is better.
+{ .jc-caption }
+
+<div class="jelly-chart-notes" markdown>
+
+- \* writes from an RDFLib `Graph`, which deduplicates statements (throughput counts only the statements written)
+
+</div>
+
+--8<-- "snippets/performance/py-b-speed-size-rdf-triples.data.html"
+
+#### Throughput per dataset
+
+--8<-- "snippets/performance/py-c-heatmap-rdf.html"
+
+Throughput on each dataset, in statements/s. The last 3 datasets contain quads.
+{ .jc-caption }
+
+<div class="jelly-chart-notes" markdown>
+
+- \* writes from an RDFLib `Graph`, which deduplicates statements (throughput counts only the statements written)
+
+</div>
+
+--8<-- "snippets/performance/py-c-heatmap-rdf.data.html"
+
+--8<-- "snippets/performance/py-d-spread-rdf-triples.html"
+
+Throughput on each dataset (dots) and the geometric mean over 10 datasets (bar; 6 for RDF/XML). Log scale.
+{ .jc-caption }
+
+<div class="jelly-chart-notes" markdown>
+
+- \* writes from an RDFLib `Graph`, which deduplicates statements (throughput counts only the statements written)
+
+</div>
+
+--8<-- "snippets/performance/py-d-spread-rdf-triples.data.html"
+
+#### Serialized size
+
+--8<-- "snippets/performance/py-f-size-rdf-triples.html"
+
+Serialized size in bytes/statement, uncompressed and with zstd, geometric mean over 10 datasets (6 for RDF/XML). Log scale.
+{ .jc-caption }
+
+<div class="jelly-chart-notes" markdown>
+
+- \* writes from an RDFLib `Graph`, which deduplicates statements (throughput counts only the statements written)
+
+</div>
+
+--8<-- "snippets/performance/py-f-size-rdf-triples.data.html"
+
 ## See also
 
 - [Jelly-SPARQL benchmarks](sparql.md)
-- [Benchmark code](https://github.com/Jelly-RDF/jelly-jvm/tree/main/jmh)
+- [Jelly-JVM benchmark code](https://github.com/Jelly-RDF/jelly-jvm/tree/main/jmh)
 - [RiverBench benchmark suite](https://w3id.org/riverbench/)
 - [Jelly-JVM]({{ jvm_link() }}) – the Jelly implementation used in the benchmarks
 - [User guide](../user-guide.md)
+- [pyjelly benchmark code](https://github.com/Jelly-RDF/pyjelly/tree/main/benchmarks)
+- [pyjelly]({{ python_link() }}) – the Jelly implementation for Python