sayrer.com

What's in a Tweet?

2025-12-27T00:00:00.000Z

Once upon a time, I worked at Twitter (it was fun back then). One unsatisfying part of that software was Tweet text processing, which was defined by a nest of regular expressions that sometimes needed to be applied in a certain order. These regular expressions were implemented in at least four languages: Ruby, Java, JavaScript, and Objective-C. To keep these languages in sync, other Twitter engineers wisely wrote a single test corpus in YAML. That was a lot of work, but it does preserve interoperability most of the time.

Twitter HQ. March 2, 2015.

What are the rules for hashtags, @mentions, URLs, replies, tweet length, and all that? It’s way more complicated than one might expect. The first step was writing a Pest grammar for use with Rust, and attempting to match the Twitter conformance suite. Without that test suite, matching the Twitter behavior would have been very difficult.

The first comment in the grammar is: “A tweet is composed of a sequence of text and entities.”.

The next comment says: “An ‘entity’ represents some special content embedded in the tweet. This grammar only covers those that have representations in the tweet text. There are other types, such as polls, that do not appear in the text.”

“Pest uses a Parsing Expression Grammar (PEG), so the order of the choices here is significant. In particular, placing URLs before other entities helps resolve ambiguities where URL fragments might also be interpreted as hashtags.

The original implementations in Ruby and the other languages used a lot of regular expressions. As a result, there are mitigations for problems like Regular expression Denial of Service - ReDoS. There are also some workarounds where regular expressions aren’t quite the right fit for the task. For example, there are only two levels of nested parentheses allowed in Tweet URLs.

In Ruby 2.6, the one that ships with macOS 26 (Tahoe), the problems with backtracking in regular expressions detailed by Russ Cox in his article Regular Expression Matching Can Be Simple And Fast (but is slow in Java, Perl, PHP, Python, Ruby, …) are easy to observe. At least for some obvious test cases, these issues are much better in Ruby 3.2 and later. They seem to have applied the memoization techniques mentioned in the article. Twitter was written with Ruby 1.8.7, so Tweets are designed to avoid most of these pitfalls.

Parsing Expression Grammar

The “entity” production is defined as:

entity =
  _{
     possible_url |
     possible_hashtag |
     possible_mention |
     possible_cashtag
  }

The ”_” underscore character is what Pest calls a silent rule. That means the grammar checks them, but they will not show up in the parse tree. These “possible” matches are there because all of these entities have characters that are not allowed to precede them, and other characters that are not allowed to follow them. The Pest grammar inspects tokens until there is a certain match. For example, “@@sayrer” might be a mention, but it’s ruled out because mentions can’t contain ”@”, and cannot be preceded by ”@”. Writing “@sayrer” will match.

Other entities, like cashtags and hashtags, have similar constraints. URLs are a bit more complex. With URLs that begin with “http://” or “https://”, the main issue in English text are strings along the lines of “https://google.com/.”. That is a valid URL, but the period is part of the text, not the URL. That trailing period should not be linked or highlighted: “https://google.com/.”. There is a further twist, though. Tweets also highlight text like “google.com”. There’s no URL scheme to start off the link there, so Tweet parsers must examine possible URLs by probing the text for potential URLs until they find a valid TLD (top-level domain) followed by characters that are allowed. This rule means text like “anything.google.com” or “anything.google.co.uk” will be linked and highlighted, but “google.com.notatld” will not.

Sometimes, these rules produce unexpected results, as in this incident. The user meant to type ”…G-20. In…”, but missed a space, so “G-20.In” was linked and highlighted. It was an unregistered domain, so someone registered it and then used it for heckling.

In addition to a long list of TLDs, the Pest grammar also includes a long rule for matching emoji. These need to be distinct from text, because emoji composed from combining characters must only count as one emoji toward the character limit.

Performance

This project didn’t start out with the aim of building a really fast Tweet parser, but its performance is monitored. The Pest grammar is very clear, but as shown below, it has some performance issues. It should be at least as fast as the Ruby and Java implementations, where most of the time is spent in their regex engines.

This section contains some language comparisons, and all of the usual benchmarking hazards are present. They might not be measuring quite the same thing, some benchmarks might be returning incorrect results, and maybe a language could be faster with a different approach. All of the tests run from the same data, and all of the conformance tests pass for all implementations. But feel free to submit a pull request.

Twitter Text: Rust/Pest vs Ruby vs Java (ops/sec, higher is better)

Rust with Pest

Ruby

Java

It seemed like the Rust code could be a lot faster. Profiles showed that the TLD and emoji rules were the most time-consuming parts of the grammar, even though they were hand-tuned a bit. Given that there was an extensive test suite, and a clear grammar, Claude Code seemed like a very good tool for writing optimizations. The first improvement attempted with Claude was to use some procedural data structures for the TLDs and emoji rules. Claude suggested adapting the code to use the emojis crate (”✨ Lookup emoji in O(1) time”), and similarly use its underlying phf crate (perfect hash functions) for the TLD list. Reasonable.

Claude Code got things mostly right. It vastly reduced the size of the grammar by using procedural code for the TLDs and emoji rules. There was already procedural code for other reasons anyway, so it adjusted the boundary of the grammar and that code. Its code needed a few tweaks, but less than ten or so. Claude Code was also strangely dismissive of failing tests and would sometimes give up, which was odd.

Twitter Text: Rust vs Ruby vs Java (ops/sec, higher is better)

Rust with PHF Emoji and TLDs

Ruby

Java

That performance is much-improved, but it’s still slower than Java in one test. JavaScript would certainly be faster still, even though it hadn’t been tested it yet. The V8 folks have been hacking and optimizing that problem for a long time. All tests had been running in both versions of the Rust implementation, and Claude Code suggested moving more code from the grammar to the procedural code.

Another way to go would have been to use regular expressions in Rust. That would probably perform pretty well. It would require the fancy-regex crate, since Tweet text uses some lookahead assertions, and the normal regex crate (intentionally) does not support these. It would also still need some of the uglier and inefficient parts of the regex-based implementations, like removing overlapping entities in post-processing.

Nom

Since it would be possible to create a procedural parser by following the Pest grammar, Claude Code attempted to read the Pest grammar and write a Nom parser. It got all of the easier parts correct, but many complex tests were still failing. With some back-and-forth instruction, Claude Code was able to get everything working. In hindsight, it probably would have been faster to write these fixes by hand.

The Nom parser was profiled in a few ways. The first step was to use samply, an easy-to-use profiler that produces flame graphs viewable in the browser. The workflow was to read the profiles and talk over any hotspots with Claude Code. It would suggest optimizations, some of which were quite bogus, but it was possible to pick and choose. Then, Claude Code would write some microoptimizations, fail tests, and then fix them. Sometimes, the suggested optimization would end up being slower by the time it was correct. A lot of them made it through, though. The biggest wins were from switching to memchr where possible, and avoiding allocations. Not coincidentally, the memchr crate is written by the author of the Rust regex crate.

At some point, samply stopped finding new things to optimize. Gungraun wraps Valgrind, which counts instructions and many other metrics that don’t rely on wall-clock time. Gungraun was able to highlight some more optimizations that were not visible in the samply flame graphs.

The downside to all of these microoptimizations is that the code is not exactly fun to read. It is not what the parser combinator folks that wrote Nom had in mind. Or, as Claude.ai put it: “End result: a hand-rolled recursive descent parser wearing a Nom trenchcoat.” or “…and suddenly you’ve got something that’s 3x faster but looks like it was written by someone who hates you.”

Well, it is faster. The following charts change the y-axis from 2,500 to 160,000. That is a really big difference, but the benchmarks are simple. Lies, damned lies, statistics, and benchmarks. However, the results are consistent across machines and in many languages. Claude Code didn’t suggest any optimizations that were obviously cheating, and all optimizations were in the Tweet code, not the test harness. They all focused on the trickier parts of the Tweet grammar.

Twitter Text: Rust vs Ruby vs Java (ops/sec, higher is better)

Rust with Nom

Ruby

Java

Those results are extremely good. Again, these benchmarks are simple. But this comparison has been run on other data sets, like large amounts of tweets, artificially long tweets, international languages, etc. The ugly Nom implementation is always at least an order of magnitude faster than the alternatives, but the interface is the same as the pure Pest version, and they are tested against each other.

Other Languages via FFI

Since the interface to the Rust library is so simple, it is easy to use from other languages via FFI. Foreign Function Interfaces (FFI) is the generic term. Each language has unique ways of calling the library, which are detailed below. The y-axes of these charts are scaled to 160k to match the Rust/Nom chart above.

Twitter Text: Java Rust FFI vs Java (ops/sec, higher is better)

Rust FFI

Java

The Java FFI implementation is accomplished via the Java Foreign Function and Memory API (FFM API) and jextract.

There’s obviously a slowdown vs the Rust/Nom version, but the code is shared and it’s still pretty fast. Regular expressions in Java are not particularly fast compared to normal Java code, so the pure Java version would probably be faster with a hand-rolled parser that didn’t cross an FFI boundary. This wouldn’t necessarily be true for some of the other languages, since the regex engine is often faster than the language itself.

Twitter Text: Ruby Rust FFI vs Ruby (ops/sec, higher is better)

Rust FFI

Ruby

The Ruby FFI implementation is accomplished via magnus. The Ruby FFI implementation is a little faster than the Java one. This shows that most of the time is spent in Rust, but the FFI boundary is a bit more expensive in the Java version. It’s probably string conversion (UTF-8 to UTF-16) in this case, but it hasn’t been measured.

This test was performed with Ruby 3.4.7. The original Twitter implementation used Ruby Enterprise Edition 1.8.7 and later. Ruby 3.4.7 is about 1.2-1.5x faster than REE 1.8.7 on the older ruby benchmarks for twitter-text 1.0.

Twitter Text: Obj-C Rust FFI vs Obj-C (ops/sec, higher is better)

Rust FFI

Obj-C

Tweets need to be processed on Apple clients, so there’s an Objective-C library. There’s an FFI directory that wraps the Rust implementation to make them callable from C (or Objective-C or C++). Rust has solid support for this task.

The original Objective-C implementation doesn’t have an Autolinker, because that’s a server side task. The Java implementation is used both on the server and in the Android client, where the Autolinker presumably goes unused.

Twitter Text: JS Rust WASM vs JavaScript (ops/sec, higher is better)

Rust WASM

JavaScript

This JavaScript/WASM test was profiled many times, and Claude Code diagnosed performance problems with a few truisms (it’s the WASM boundary etc). Everything it mentioned was in the profile, but none of it really dominated. The slowdown vs the native Rust implementation is large, but it seems to be just WASM itself that is slower. That’s expected, but it seems to be really costly in this test case.

This is quite a good result, given that the competition is V8 and its highly optimized regex JIT that generates native code.

New Languages

It’s possible to add other languages that haven’t previously had a Tweet parser.

Twitter Text: Rust vs C++ vs Swift vs Python (ops/sec, higher is better)

Rust

C++

Swift

Python

This chart shows the result of adding C++ (as well as an ASAN build), Swift on macOS/Linux, and Python through PyO3.

Mastodon

Another thing that’s easier to do now is add new syntax across all of these implementations. Mastodon, a decentralized social network, happens to use twitter-text for highlighting text inside posts. They omit some features, like cashtags, and add their own @mention name highlighting.

Their mention syntax is something like @user@domain.tld. Support for this syntax was added in one commit across all languages. It was not a difficult change, but it was accomplished all at once.

Bazel

How is this all held together? Bazel. The MODULE.bazel file has all of the dependencies for the project. Rust, Java, C++/LLVM, Python, Swift, etc. Once Bazel is working, it will download all of that and start building.

There is one exception to that hermetic setup. The Bazel Ruby rules don’t work well with Bazel’s remote execution, so those require a local Ruby installation, and the Ruby tests don’t run in the remote build configuration.

The kind folks at BuildBuddy run all of the tests remotely. Here’s an example build log. So, on every commit, BuildBuddy runs the tests, but it only builds and tests what has changed. GitHub Actions are used to trigger rustfmt and clippy.

Cross Platform Applications

Building everything as a light wrapper around Rust is possible, but it doesn’t make sense if an application uses a lot of platform-native code (detailed UI, cameras, motion sensors, etc). In those cases, trying to do everything in Rust just makes the task more difficult, and probably duplicates code anyway. OpenAI has a case study where they used LLMs for their Sora Android port. For now, there does seem to be a sweet spot where server and client side logic can be written in Rust or equivalent, while porting the UI quickly to a native platform with LLMs.

This project should also serve as a recipe for a fairly complex polyglot monorepo.

Istanbul, Türkiye

2019-06-15T00:00:00.000Z

Istanbul, Türkiye. June 15, 2019.

I thought I took more pictures, but maybe they were just Instagram Stories or something. This is from the one ferry trip I took. It went to the Black Sea and back. I used the streetcars or walked most of the time. The only taxi I used was to/from the airport, which is far away and huge.

This city is highly recommended.

Hawaiʻi Island

2019-06-03T00:00:00.000Z

Hawaiʻi Island. June 3, 2019.

Hawaiʻi Island. June 4, 2019.

Taipei, Taiwan

2019-04-27T00:00:00.000Z

Taipei, Taiwan. April 27, 2019.

Mexico City, Mexico

2019-04-08T00:00:00.000Z

Mexico City, Mexico. April 8, 2019.

Santiago, Chile

2019-02-08T00:00:00.000Z

Santiago, Chile. February 8, 2019.

Difficult to complain about that view. The one thing there was time for was a visit to Pablo Naruda’s house.

Santiago, Chile. February 9, 2019.

Valparaíso and Ritoque

2019-02-06T00:00:00.000Z

Valparaíso, Chile. February 6, 2019.

After Valparaíso, we visited Ritoque. It was a socialist vacation resort under Salvador Allende. It still has that vibe, even though the cabins are gone. It’s a people’s beach.

Ritoque, Chile. February 8, 2019.

Ritoque, Chile. February 9, 2019.

We told our host in Ritoque that we were planning to visit Ciudad Abierta (Open City). He told us that many cars get broken into there, so to park at the neighboring Go-Kart track and walk up the train tracks.

Ciudad Abierta, Chile. February 11, 2019.

The advice to walk up the train tracks was good, but it turns out there are actual trains.

Bahia Murta

2019-02-05T00:00:00.000Z

Bahia Murta, Chile. February 5, 2019.

In other posts, I warned to book Puerto Río Tranquilo ahead. Here’s how I know. There was nothing except for this Airbnb llama farm a few km up the road at Bahia Murta. It ended up being excellent, but it sure sounded like “Muerta” to me.

At this place, the road is so curvy that you must honk the horn as you go around the corners, so people know you’re coming.

As it happens, there is a nearby attraction called Cascada Oculta de Bahia Murta.

Bahia Murta, Chile. February 5, 2019.

After this, we gave two of the Airbnb hosts a ride on the way out. We went to Balmaceda Airport, and exited Chilean Patagonia.

Chile Chico

2019-02-04T00:00:00.000Z

Puerto Guadal, Chile. February 4, 2019.

Route 265, Chile. February 4, 2019.

Chile Chico, Chile. February 4, 2019.

This drive is taxing. Be prepared.

Parque Nacional Patagonia

2019-02-03T00:00:00.000Z

Parque Nacional Patagonia, Chile. February 3, 2019.

So… it turns out you need to get a permit for crossing into Argentina when you rent the car. We had to turn around. The border guard was super nice and tried to get them to give me a pass over the radio. We didn’t get an answer, but I got to see all of these cool old radios inside the building.

Carretera Austral Part 8: Cochrane to Villa O'Higgins

2019-01-31T00:00:00.000Z

Puerto Yungay, Chile. January 31, 2019.

Yes, there are still Pinochet signs in a few places.

Villa O'Higgins, Chile. January 31, 2019.

The end of the road. Villa O’Higgins.

Carretera Austral Part 7: Lago General Carrera to Cochrane

2019-01-30T00:00:00.000Z

Puerto Bertrand, Chile. January 30, 2019.

Rio Baker, Chile. January 30, 2019.

In a traffic jam outside of Cochrane, Chile. January 30, 2019.

Carretera Austral Part 6: Coyhaique to Lago General Carrera

2019-01-29T00:00:00.000Z

Cerro Castillo, Chile. January 29, 2019.

Puerto Río Tranquilo, Chile. January 29, 2019.

Puerto Río Tranquilo is one of those places you need to plan ahead for. This is really popular for family rafting trips and things, as shown.

Lago General Carrera, Chile. January 29, 2019.

Then, we crossed a bridge, and I stopped to take a picture.

Puente General Carrera, Chile. January 29, 2019.

After that, we checked in to Pared Sur Camp. We ended up in a regular room, not one of the geodesic domes, but it was good.

Carretera Austral Part 5: Puyuhuapi to Puerto Aysén to Coyhaique

2019-01-26T00:00:00.000Z

Cascada de Ventisquero Colgante, Chile. January 26, 2019.

This distance shot is of Cascada de Ventisquero Colgante in Queulat National Park. It’s the main attraction along this route.

Approaching Cascada de Ventisquero Colgante, Chile. January 25, 2019.

Cascada de Ventisquero Colgante, Chile. January 26, 2019.

That is indeed a waterfall falling off a glacier.

Cascada de Ventisquero Colgante, Chile. January 26, 2019.

OK, all done. A somewhat steep hike, but short. One thing you have to do on the Carretera Austral is help college student hitchhikers as a matter of civic duty, and there is only really one road. We did that here, since we were headed in the same direction, and gave a young couple a ride on the way to Puerto Aysén.

Little did I know that there were absolutely crushing switchbacks to follow. They start at Salto Padre García. Zoom in on the map and laugh at my ignorance, and keep in mind that the incline is steep at all points.

So, there are no pictures of this part.

Puerto Aysén, Chile. January 26, 2019.

So, Puerto Aysén is all business. A port town. Still not bad looking, though.

We stayed at Hotel y Cabañas Patagonia Green, which was fine. Quiet and convenient, but you have to drive over a bridge from downtown, which is a little confusing when you’re tired.

Finally, we arrive at Coyhaique.

Coyhaique, Chile. January 27, 2019.

This is the view from the second (or first, depends on your localization) floor of Mamma Gaucha.

If you’ve gotten this far, you might be asking about good coffee. You can probably find it in Coyhaique, but the shops are always changing. You can also do laundry easily—we just did a wash-n-fold here.

If I could do it again, I would spend more time between Caleta Gonzalo and Coyhaique.

Carretera Austral Part 4: Lago Yecho to Puyuhuapi

2019-01-25T00:00:00.000Z

La Junta, Chile. January 25, 2019.

Puyuhuapi, Chile. January 25, 2019.

Puyuhuapi dogs running free in the town.January 25, 2019.

The owner in the white pickup truck has come home.

Carretera Austral Part 3: Caleta Gonzalo to Lago Yecho

2019-01-24T00:00:00.000Z

Leaving Caleta Gonzalo, Chile. January 23, 2019.

I stopped to take this picture because it looked like Indiana Jones or something.

There aren’t many pictures from this part of the trip, because the driving was a little tough. The roads weren’t too bad, but there were lots of those big trucks you might have seen from the ferry pictures.

Chaitén, Chile. January 23, 2019.

Lago Yecho, Chile. January 23, 2019.

We stayed at Hotel Yelcho en la Patagonia. I didn’t take any pictures of it, but the ones above are from their yard. I think I was just tired. They had an excellent gift shop full of Patagonia-branded items I’d never seen. The serious fishing gear.

Carretera Austral Part 2: Hornopirén to Caleta Gonzalo

2019-01-23T00:00:00.000Z

Hornopirén, Chile. January 23, 2019.

The guidebook will tell you book this one far in advance. It’s with Somarco. The problem is that you need to provide your license plate, while you must book it before you have your rental car. You can just put “NNNNNN” and go to the ferry office by the pier and change it when you arrive. At least, that’s what I did.

Hornopirén to Caleta Gonzalo, Chile. January 23, 2019.

Arriving to Caleta Gonzalo, Chile. January 23, 2019.

I stayed in the little house on the right. Caleta Gonzalo is pretty fancy, so also book this ahead. But you can camp without reservations in a big yard, too.

Caleta Gonzalo, Chile. January 23, 2019.

Went for a hike the next morning. It was a little tougher than expected.

Caleta Gonzalo, Chile. January 24, 2019.

Carretera Austral Part 1: Puerto Montt to Hornopirén

2019-01-22T00:00:00.000Z

Puerto Montt, Chile. January 21, 2019.

The Carretera Austral starts in Puerto Montt. It’s nice. It reminded me of one of those medium-sized towns in the Pacific Northwest. A mix of wilderness and a few ugly parts on the outskirts.

Waterfront. Puerto Montt, Chile. January 21, 2019.

The drive starts in the city easy enough. But you hit pretty rough road construction pretty quickly. Gravel roads and all that. Fortunately, we had a Toyota 4Runner (which is the right choice), and it was pretty easy. It’s pictured below. This where you hit the first ferry of the trip. It’s short, and starts at Caleta La Arena.

Conecta La Arena con Puelche, Chile. January 21, 2019.

Every ferry company is different. This one is Transportes del Estuario.

After that, it’s a short drive to Hornopirén. The trip takes a while, because you have to wait for the ferry.

Here’s the downtown:

Hornopirén, Chile. January 21, 2019.

Other parts look like this:

Hornopirén, Chile. January 21, 2019.

Dogs in Chile just roam around the town and everyone takes care of them. This one decided he wanted to follow me to see tomorrow’s boat at low tide.

Carretera Austral

2019-01-21T00:00:00.000Z

The Carretera Austral guidebook. January 21, 2019.

The Carretera Austral is a road that runs 770 miles through Chilean Patagonia.

Santiago, Chile

2019-01-20T00:00:00.000Z

Santiago, Chile. January 20, 2019.

The first two images are of Mercado Central de Santiago.

Santiago, Chile. January 20, 2019.

This the view from Cerro Santa Lucía.

Santiago, Chile. January 20, 2019.

This is the view from below Cerro Santa Lucía.

Easter Island

2019-01-13T00:00:00.000Z

Moai, Easter Island, January 15, 2019.

Might as well start with the expected ones. You can walk almost right up to them, but there’s a narrow path you must stay on. Otherwise, someone will blow a whistle and yell at you. I was not scolded, and most people were well-behaved. It was usually just people backing up while taking a photo.

Half-carved Moai, Easter Island, January 15, 2019.

Ahu Tongariki, the famous row of Moai. Easter Island, January 15, 2019.

Ahu Tongariki

Ahu Ature Huki, Moai statue standing on the hill at Anakena. Easter Island, January 15, 2019.

Anakena Beach, Easter Island, January 15, 2019.

Anakena Beach is said to be the first place the initial Polynesian people arrived.

Farm fields. Easter Island, January 16, 2019.

Ahu Vai Ure, Easter Island, January 16, 2019

This is Ahu Vai Ure, right near Hanga Roa, the town.

Izakaya Kotaro / Kotaro bistro, January 16, 2019.

This is an Izakaya that the owner constructed himself. It’s here.

Izakaya Kotaro / Kotaro bistro, January 16, 2019.

Rano Kau, Easter Island, January 20, 2019.

Moto Nui and Motu Iti, Easter Island, January 20, 2019.

Singapore

2018-12-05T00:00:00.000Z

Andaz Singapore, December 5, 2018.

Building across the street from the hotel, Singapore, December 5, 2018.

Building across the street from the hotel from the room, Singapore, December 5, 2018.

But what’s the neighboorhood down there? Let’s see.

Street, Singapore, December 5, 2018.

Sultan Mosque, Singapore, December 5, 2018.

Arab Street, Singapore, December 5, 2018.

The Great Wall of China

2018-11-04T00:00:00.000Z

The Great Wall of China, November 4, 2018.

This last one is in a different section, where the wall isn’t in such great shape.

The Great Wall of China, November 4, 2018.

Beijing, China

2018-10-27T00:00:00.000Z

Beijing, China October 27, 2018.

Beijing subway, China October 28, 2018.

I took the subway to…

Tiananmen square, Beijing, October 28, 2018.

I didn’t take any pictures inside the Forbidden City, because they would just be the same as everyone else’s. I did take one picture on the road after the exit, though.

Forbidden City exit street, Beijing, October 28, 2018.

Later on, I got to check out some modern consumerist areas.

Acne Studios, Beijing, November 4, 2018.

OK, they got me with this one:

Adidas, Beijing, November 4, 2018.

Lisbon, Portugal

2018-09-26T00:00:00.000Z

Lisbon, Portugal. September 26, 2018.

The view from Castelo de São Jorge. If you’re looking for good coffee not in the espresso Southern European style, I can recommend Copenhagen Coffee Lab. They have have many locations, including one right near here. Alfama.

Tunis, Tunisia

2018-09-13T00:00:00.000Z

In case there’s confusion, I am not Tunisian. But I have gone to school there, and have been visiting for years. You’re getting just about the best advice you can get if your world view is from the west coast of the United States.

Tunis, Tunisia. September 12, 2018.

It’s Bab al-Bhar (The gate to sea) pictured above. They say the modern city is mostly landfill. Maybe the gate wasn’t quite on the beach in the old days, but it’s tough to tell. It does lead in that direction.

For all of the fancy things, the guidebooks are generally correct. The expensive restaurants in the Medina are good. These pictures are from Hotel Royal Victoria. This used to be the British Consulate. The picture above is from a room, and the video below is the elevator.

Hotel Royal Victoria. September 13, 2018.

This hotel is the best choice for your first visit. There are plenty of other cool places, and be aware you will be in the mix right away. If you’ve never been to North Africa or the Middle East, it might be a lot. The café downstairs is easy to deal with. It will go slow, but it’s ok to flag down the waiter. They try not to bother you.

I will not publish the cool places I go, but I will tell you about them if you use the contact page.

Algiers to Tunis

2018-09-11T00:00:00.000Z

There is no iOS app.

TunisAir September 11, 2018.

Flutter really works well in these markets. That way, the Android app is good, and the iOS app works.

Here “vol” means “flight”. So, it’s “book a flight”, “my flight”, “flight status”. Lastly, it’s “check in”.

Algiers, Algeria

2018-09-10T00:00:00.000Z

My photos don’t do the city justice. It is beautiful, but I have an excuse.

El Aurassi Hotel. Algiers, Algeria. September 9, 2018.

El Aurassi Hotel is a pretty fancy hotel for a centrally located one.

I sat at the hotel restaurant on my first evening. I was kind of tired and crabby, so I sat in the back next to a vacant grand piano. The restaurant is an ellipse. I quickly discovered that I could hear a whisper at the restaurant bar from 100ft (~30m) away.

Algiers, Algeria. September 9, 2018.

Algiers has a lot of steep hills, so you see a lot of views like this one.

Algiers, Algeria. September 9, 2018.

I was hiking around the medina/casbah (the old city). I got to the top of the hill, and there were a bunch of government buildings up there. A policeman started talking to me in Arabic.

Everyone in Algeria assumed I would understand Arabic since I wasn’t in a tour bus and didn’t look lost. In big cities, the signs are in French as well as Arabic, and I can read the really basic Arabic ones.

When he found out I was from the US, he did a doubletake. He used French, and told me to go back down the hill.

By the time I got back down to the bottom of the hill, I noticed there were two shady guys in bad suits following me. It probably seemed like I was trying to detect them by walking with no direction. I did see one of them peer out from behind a fruit cart. So, I stopped taking pictures.

Oran to Algiers

2018-09-09T00:00:00.000Z

No images in this one. The reason will become clear.

If you know me, you’ll know that I mostly dress in black. Nothing fancy, just H&M or Uniqlo. Maybe some Burton, Nike, or Hurley for the outerwear.

But what you probably don’t know (I didn’t) is that the airport in Oran, Algeria has two terminals. One terminal is for international flights, and the other is for domestic flights. So, if you’re flying from Oran to Algiers, you’ll use the domestic terminal.

So, I’m just wearing my usual basic black clothing, with an added Raiders hat listing no city. No problem, who cares most of the time.

But what I also didn’t know is that the Hajj flights depart from the domestic terminal.

Everyone was really nice to me, clearly understanding that I was clueless. But I was dressed like a pirate amongst the pilgrims, who dress all in white.

Oran, Algeria

2018-09-07T00:00:00.000Z

Oran, Algeria. September 7, 2018.

Oran, Algeria. September 8, 2018.

So, I was up on that hill overlooking that bay. I had researched it, and there were rumors of a secret Soviet submarine base. I didn’t find evidence of it, but I did find numerous WWII fortifications and tunnels. Not too different from Battery Spencer in San Francisco.

Then I did something you should never do unless you’re really sure you can handle it. In hunting for the Soviet submarine base, I had read French text (my French is not that great, but I can read maps). So, I knew there was an ancient Roman road that switchbacked up the hill. When I looked, it was still there.

I walked down from the top of the hill on the Roman road. Google Maps had everything mislabeled, and I only knew where I was from reading the old French books. Eventually, I ended up at what was called “Ft. Saint Gregoire”, but is now called Hassan bin Zahwa Tower (Dyar El Houmrin) قلعة سان جريجويو . I sent Google Maps corrections for all of these (numerous!) wrong labels. They fixed them, but now they all have Arabic names.

That’s where I ran into a bodybuilder dude. Il a dit “Tu es seul? Tu vas en bas? …Doucement.” He meant the neighborhood down there was sketchy, and to take care. He was right, that place was dangerous, but I got through it.

Lisbon, Portugal

2018-07-28T00:00:00.000Z

Lisbon, Portugal. September 4, 2018.

Lisbon, Portugal. September 6, 2018.

That logo is Globo. It is like Uber Eats or DoorDash.Globo’s growth team is remarkably persistent. I still get Portuguese Siri messages in an English voice. So, now I know inter-English-Portuguese-AInglish.

Brooklyn, New York

2018-07-24T00:00:00.000Z

Brooklyn, New York. July 24, 2018.

This space used to be my bedroom, but now it’s a pod hotel courtyard. I stayed there just to see how it felt. You can’t go home again, as they say.

Reykjavík, Iceland

2018-07-23T00:00:00.000Z

Reykjavík, Iceland. July 19, 2018.

Reykjavík, Iceland. July 23, 2018.

So, if the jetlag is really bad, you walk out to this place. It’s farther than you think, and you take off your shoes and put them in the hotsprings. Difficulty is low, but it’s pretty far with no hills. You be the only non-Icelandic person there, unless you’re Icelandic. There are also other bootleg hotsprings nearby, that’s where the cool people are.

Honolulu, Hawaii

2018-07-19T00:00:00.000Z

Honolulu, Hawaii. July 17, 2018.

Mexico City, Mexico

2018-07-18T00:00:00.000Z

Mexico City, Mexico. June 23, 2018.

Mexico City, Mexico. June 26, 2018.

Parque Nacional Iztaccíhuatl Popocatépetl, Mexico

2018-07-12T00:00:00.000Z

Parque Nacional Iztaccíhuatl Popocatépetl, Mexico. June 26, 2018.

Parque Nacional Iztaccíhuatl Popocatépetl, Mexico. June 2018.

I got a little bit of altitude sickness from this adventure. Mexico City is already at an elevation of 7,350ft (2,240 meters), but that was not adequate preparation. We didn’t go to the summit of Iztaccíhuatl, but we did reach roughly 13,000ft (4,000 meters). I never felt disoriented or anything while I was up there, but I did feel crummy for a few days afterward.

Teotihuacan, Mexico

2018-06-16T00:00:00.000Z

Teotihuacan, Mexico. June 19, 2018.

Rio de Janeiro, Brazil

2018-05-18T00:00:00.000Z

Rio de Janeiro, Brazil. May 9, 2018.

Rio de Janeiro, Brazil. May 10, 2018.

Rio de Janeiro, Brazil. May 16, 2018.

Rio de Janeiro, Brazil. May 2018.

Rio de Janeiro, Brazil. May 16, 2018.

Salvador, Brazil

2018-05-18T00:00:00.000Z

Salvador, Brazil. May 11, 2018.

São Paulo, Brazil

2018-05-17T00:00:00.000Z

São Paulo, Brazil. May 20, 2018.