# Searching for RH Counterexamples — Scaling Up

We’re ironically searching for counterexamples to the Riemann Hypothesis.

Last time we made the audacious choice to remove primary keys from the `RiemannDivisorSums` table for performance reasons. To help with that, we will do two things in this post

• Reduce the storage footprint of the whole application (it was 60 GiB when it crashed, and we got up to 84 prime factors).
• Refactor the application into a worker architecture.

The first one is straightforward, but uses a possibly new idea to the mathematician, and the second is a relatively major rearchitecture.

## Hashing a search block

Instead of storing every single Riemann divisor sum, we can just store the ones that are interesting, that is, the ones that are larger than some threshold—in our case, 1.767, which had just under a thousand examples in the last post.

That raises a problem. We want to be able to say that we checked all examples up to some limit. We’ve been saying “up to all numbers having at most K prime factors” because that’s how the superabundant search strategy enumerates the candidates. If we don’t store all the witness values, how can we prove to a skeptical observer that we checked what we said we checked?

One approach is to instead store a “summary string” derived from the full set of witness values. This “summary string” should be deterministically reproducible, so that if someone has the metadata about the block of numbers, they can recompute the summary and compare it with our summary.

A simple example summary might be the string-concatenation of all the witness values one after another, like `1.4323,1.5678,0.8792`. But that still takes up a lot of space (each block had 250,000 different witness values). It would be better if we could find a summary that had a fixed size, regardless of how big the input is. Taking a cue from computer science, a hash function does exactly that. Hash functions aren’t perfect, they can have collisions, but the collisions are unlikely enough that, if you tested a hundred random blocks and all your hashes agreed with all of my hashes, you would be confident I checked everything the same as you. That gives you confidence that the rest of my blocks (say, a million of them) were computed correctly.

The hash function we’ll use is called sha256, and we add it in this pull request. Now the database application shouldn’t need a ton of disk space to run. (We’ll confirm this later in the article)

For more background and related topics around hash functions, see Load Balancing and the Power of Hashing and Hashing to Estimate the Size of a Stream.

## Worker architecture

Next we want to rearchitect the application to enable two things: better protection against the possibility of duplicating/skipping divisor sum computations, and allowing us to scale up the rate of computation by adding additional computers to the mix. This pull request achieves the goal, but note that it’s a large change, so let’s cover the broad ideas, then zoom in on a few details, and finally discuss the PR’s structure as a whole from the perspective of software engineering.

We start by adding a “state” field to a search block. The state can be one of `NOT_STARTED, IN_PROGRESS, FINISHED, FAILED`. Then one worker job (read, container) will be responsible for computing new blocks in the `NOT_STARTED` state, and a group of workers jobs will continually “claim” an eligible block, compute the divisor sums, and then mark it as finished and insert the interesting witness values it found.

This suggests a drastic change to the database interface, as shown in this commit, to align with the new “claim/finish” workflow. We replace the existing functions with `insert_search_blocks`, `claim_next_search_block`, and `finish_search_block`. The first two operate only on the `SearchMetadata` table, while the third couples together the changes to both tables, by marking a block as finished and inserting the relevant witness values. We’ll see how we ensure these two updates occur atomically in a moment. Also, because of the new coupling of the two tables, I merged the two database interfaces into one.

We need two safety measures to make this work. First, we need to ensure that we don’t accidentally generate multiple duplicate search blocks. We do this by adding a uniqueness constraint on the bounds of a search block at the database level. If a client calls `insert_search_blocks` with something that’s already in the database, the database itself will reject it, and the client will be forced to catch the failure and recover.

Second, we need to ensure that two workers don’t claim the same search block. This involves a bit of database wizardry. I had to do a bit of research to make it work, and it took a few attempts. The gist of the problem is that if you want to claim a search block, you need to do two queries. First, you need to look up the oldest eligible search block that you want to claim. Then, you need to mark it as claimed. But if you have multiple workers accessing the database simultaneously, you can get this order of operations

1. Worker 1 runs lookup query, gets search block 7
2. Worker 2 runs lookup query, also gets search block 7
3. Worker 1 runs update query, marks search block 7 as claimed
4. Worker 2 runs update query, also marks search block 7 as claimed.

This can even happen if there’s only “one” query, and the lookup step is a subquery. After reading a bit about PostgreSQL’s transaction isolation guarantees, I learned that you can “lock” the rows read by the subquery until the update step is complete by adding `FOR UPDATE` to the subquery. This commit has the full query, but it looks something like this

```UPDATE SearchMetadata
SET state = 'IN_PROGRESS'
FROM (
SELECT starting_search_index
WHERE state = 'NOT_STARTED'
ORDER BY creation_time ASC
LIMIT 1
FOR UPDATE         <---- key clause!
) as m
WHERE
starting_search_index = m.starting_search_index;
```

I also added some special tests (that use the finished worker jobs) that will specifically fail when this `FOR UPDATE` clause is removed, to confirm it fixes the behavior. This is particularly important because it’s hard to test database queries that protect against race conditions. Plus, I tried other approaches and got it wrong, which proved the value of having such tests.

Finally, the `finish_search_block` function needs to ensure that the update of the `SearchMetadata` and `RiemannDivisorSums` tables happen atomically. This is done in this commit by leveraging the concept of a database transaction. In a transaction, all of the queries between `BEGIN` and `COMMIT` happen at once or not at all, e.g., in the case that a later query fails. The “not at all” case is called a rollback.

Thankfully, the `psycopg2` library we’re using uses transactions by default, and requires us to call `connection.commit` to make any query persist. So we are already using transactions and get the guarantee without extra work.

Next, we needed to update the `SearchStrategy` interface to separate the task of generating new search blocks and processing blocks, since these will be done by separate workers. This commit updates the interface, this commit the tests, and this commit the implementation.

Finally, once all the rearchitecting and re-testing is done, we ditched our old `populate_database` entry script, and replaced it with a worker script for generating new search blocks, and one for claiming and processing blocks. These have the same general structure: infinitely loop, use the database and search strategy interfaces as expected, and then exponentially back off when failures occur, eventually quitting if there are too many repeated failures. In the case of the generator worker, we also have configuration around how many new blocks to create, what minimum threshold to trigger doing work, and a delay between checks (since the workers will be much slower than the generator).

## Ship it!

Now we’re ready to launch it, and this time since we have improved our memory issues, we can use an AWS instance with smaller memory requirements, but launch four of them to allow one for each container. After a few of the following steps, this worked out just fine:

• Configure an AWS security group (which limits which IP addresses EC2 instances can communicate with) so that Postgres communication was allowed between all the launched instances. This was made easy by the fact that you can configure a security group once so that it allows communication to and from any other instances with the same security group (in addition to my laptop’s IP).
• Keep track of the ip addresses of each of the instances I launched, and make sure that the instance with a 60 GiB disk is the instance I put the database on.
• Manually install docker and launch the right container on each one, and point their `PGHOST` variables to the address of the database instance.

I will improve this workflow in the next article, but for now I am letting it run for a few days and observing that it works. The generator job adds 100 new search blocks any time there are less than 100 eligible blocks, and this query shows they are getting worked on properly.

```divisor=# select state, count(*) from searchmetadata group by state;
state    | count
-------------+-------
NOT_STARTED |   129
FINISHED    |   367
IN_PROGRESS |     4
```

As I was stopping and restarting jobs, two search blocks got stuck in the “in_progress” state, as you can see from the lagging-behind start time.

```divisor=# select start_time, starting_search_index, ending_search_index from searchmetadata where state='IN_PROGRESS' order by start_time asc;
start_time         | starting_search_index | ending_search_index
----------------------------+-----------------------+---------------------
2021-02-09 04:46:13.616085 | 64,433492             | 64,683491    <-- stuck!
2021-02-09 04:48:08.554847 | 65,191862             | 65,441861    <-- stuck!
2021-02-09 14:22:08.331754 | 78,9803652            | 78,10053651
2021-02-09 14:22:44.36657  | 78,10053652           | 78,10303651
(4 rows)
```

This suggests I should create a new job for cleaning up “stale” `IN_PROGRESS` blocks and marking them as failed. I’ll do this in the next article, and it won’t hurt us to leave some blocks in the transient state, because once the cleanup job is running it will make those blocks eligible to be worked on again, and the worker’s claim the earliest block first.

I also noticed an issue where I forgot a commit statement, which didn’t show up in testing, and I can’t quite figure out how to set up a test that requires this commit statement to be present. Nevertheless, after I noticed the problem, I patched it, stopped the containers, ran `git pull` on each instance, rebuilt the containers, and started them again, and it continued along as if nothing was wrong.

Since the search application is running well now for long periods of time, I’ll let it run for a week or two (or until something breaks), and return to it in the next article to see what bigger witness values we find. In the mean time, I’d like to discuss how I organized all of the changes that went into that massive pull request.

## How to organize a big pull request

The pull request to change the application architecture is quite substantial, and I had a lot of cleanup I wanted to do along the way.

The basic pattern of the PR, which was repeated a few times throughout, was to split each unit of change into three parts: update the interface, update the tests that use the interface, and update the implementation. As an example, one commit updates the SearchStrategy interface, one the tests, and one the implementation.

Doing it this way helps in a few ways. First, I think a lot about how the interface should work, and I prove that it feels right by writing tests first. That way, I don’t get bogged down in the implementation details when thinking about how it will be used. Likewise, when doing this for the database change, after updating the interface and tests, I first updated the `InMemoryDatabase` implementation, because it was simpler, and having this made me confident the interface could be implemented decently, and allowed me to tweak the interface early, before turning to the harder problem of figuring out PostgreSQL `FOR UPDATE` stuff.

Much like the proof of a mathematical theorem, the final commits hide the work that went into this, and commit messages, code comments, and PR description provide a deeper explanation. Typically what I do is make changes, and then use git’s staging area concept to pull out different sub-changes within a file into a single commit. This blog post about “using git well” covers this in more detail. But this, and git’s rebase and amend features, allowed me to edit older commits so that the final pull request lays out the commits in a logical order.

It also allows me to split off the cleanup parts as best as possible, such as this commit to rename SearchState to SearchIndex (after I added the new “state” field). I made the commit, had some places I missed and fixed later, extracted them into their own commit, and rebased and squashed them together to unify the change. Keeping each commit as close to an atomic thought as possible (“add this test” or “implement this method”) enables easy squashing and reordering.

Though nobody is reviewing my code now (except you, dear reader, after the fact), a primary benefit of all this cleanup is that the code is much easier to review. The pull request as a whole is a feature, each commit is a smaller, but comprehensible piece of that bigger feature, and the commits are laid out in such a way that if you browse the commits in order, you get a clear picture of how the whole feature was assembled.

If you find yourself doing mathematical work in the context of software, you should expect to spend a lot of time reviewing code and having your code reviewed. The extra effort you put into making that process easier pays off by producing better feedback, uncovering more bugs, and reducing total review time. Though it might be considered a soft skill, smooth code reviews are a critical part of a well-functioning software organization.

# Searching for RH Counterexamples — Performance Profiling

We’re ironically searching for counterexamples to the Riemann Hypothesis.

In the last article we ran into some performance issues with our deployed docker application. In this article we’ll dig in to see what happened, fix the problem, run into another problem, fix it, and run the search until we rule out RH witness-value-based counterexamples among all numbers with fewer than 85 prime factors.

When debugging any issue, there are some straightforward steps to follow.

1. Gather information to determine where the problem is.
2. Narrow down the source of the problem.
3. Reproduce the problem in an isolated environment (on your local machine or a separate EC2 instance).
4. Fix the problem.

So far what I know is that the search ran for days on EC2 with docker, and didn’t make significantly more progress than when I ran it on my local machine for a few hours.

## Gathering information

Gathering information first is an important step that’s easy to skip when you think you know the problem. In my experience, spending an extra 5-10 minutes just looking around can save hours of work you might spend fixing the wrong problem.

So we know the application is slow. But there are many different kinds of slow. We could be using up all the CPU, running low on free RAM, throttled by network speeds, and more. To narrow down which of these might be the problem, we can look at the server’s resource usage.

The CPU utilization EC2 dashboard below shows that after just a few hours of running the application, the average CPU utilization drops to 5-10% and stays at that level. Running `docker stats` shows that the search container has over 5 GiB of unused RAM. `df -h` shows that the server has more than 60% of its disk space free.

The fact that the CPU spikes like this suggests that the program is doing its work, just with gaps in between the real work, and some sort of waiting consuming the rest of the time.

Another angle we can look at is the timestamps recorded in the SearchMetadata table.

```divisor=# select * from searchmetadata order by start_time desc limit 9;
start_time         |          end_time          |       search_state_type       | starting_search_state | ending_search_state
----------------------------+----------------------------+-------------------------------+-----------------------+---------------------
2021-01-01 15:40:42.374536 | 2021-01-01 16:15:34.838774 | SuperabundantEnumerationIndex | 71,1696047            | 71,1946047
2021-01-01 15:06:13.947216 | 2021-01-01 15:40:42.313078 | SuperabundantEnumerationIndex | 71,1446047            | 71,1696047
2021-01-01 14:32:14.692185 | 2021-01-01 15:06:13.880209 | SuperabundantEnumerationIndex | 71,1196047            | 71,1446047
2021-01-01 13:57:39.725843 | 2021-01-01 14:32:14.635433 | SuperabundantEnumerationIndex | 71,946047             | 71,1196047
2021-01-01 13:25:53.376243 | 2021-01-01 13:57:39.615891 | SuperabundantEnumerationIndex | 71,696047             | 71,946047
2021-01-01 12:59:14.58666  | 2021-01-01 13:25:53.331857 | SuperabundantEnumerationIndex | 71,446047             | 71,696047
2021-01-01 12:43:08.503441 | 2021-01-01 12:59:14.541995 | SuperabundantEnumerationIndex | 71,196047             | 71,446047
2021-01-01 12:27:49.698012 | 2021-01-01 12:43:08.450301 | SuperabundantEnumerationIndex | 70,4034015            | 71,196047
2021-01-01 12:14:44.970486 | 2021-01-01 12:27:49.625687 | SuperabundantEnumerationIndex | 70,3784015            | 70,4034015
```

As you can see, computing a single block of divisor sums takes over a half hour in many cases! This is nice, because we can isolate the computation of a single block on our local machine and time it.

## Narrowing Down

Generally, since we ran this application on a local machine just fine for longer than we ran it in docker on EC2, I would think the culprit is related to the new thing that was introduced: docker and/or EC2. My local machine is a Mac, and the EC2 machine was Ubuntu linux, so there could be a difference there. It’s also strange that the system only started to slow down after about two hours, instead of being slow the whole time. That suggests there is something wrong with scaling, i.e., what happens as the application starts to grow in its resource usage.

Just to rule out the possibility that it’s a problem with computing and storing large blocks, let’s re-test the block with starting index `71,196047` and ending index `71,446047`, which took approximated 16 minutes on EC2. These three lines are between the start/end timestamps in the table. Only the second line does substantive work.

```start_state = search_strategy.search_state()
db.upsert(search_strategy.next_batch(batch_size))
end_state = search_strategy.search_state()
```

First we’ll just run the `next_batch` method, to remove the database upsert from consideration. We’ll also do it in a vanilla Python script, to remove the possibility that docker is introducing inefficiency, which is the most likely culprit, since docker is the new part from the last article. This commit has the timing test and the result shows that on average the block takes two minutes on my local machine.

Running the same code in a docker container on EC2 has the same result, which I verified by copying the `divisorsearch.Dockerfile` to a new dockerfile and replacing the entrypoint command with

```ENTRYPOINT ["python3", "-u", "-m", "timing.ec2_timing_test"]
```

Then running (on a fresh EC2 instance with docker installed and the repo cloned)

```docker build -t timing -f timing.Dockerfile .
docker run -dit --name timing --memory="15G" --env PGHOST="\$PGHOST" timing:latest
docker logs -f timing
```

Once it finishes, I see it takes on average 40 seconds to compute the block. So any prior belief that the divisor computation might be the bottleneck are now gone.

Now let’s try the call to `db.upsert`, which builds the query and sends it over the network to the database container. Testing that in isolation—by computing the block first and then timing the upsert with this commit —shows it takes about 8 seconds to update an empty database with this block.

Both of these experiments were run on EC2 in docker, so we’ve narrowed down the problem enough to say that it only happens when the system has been running for that two hour window. My best guess is something to do with having a large database, or something to do with container-to-container networking transfer rates (the docker overhead). The next step in narrowing down the problem is to set up the scenario and run performance profiling tools on the system in situ.

I spin back up an r5.large. Then I run the deploy script and wait for two hours. It started slowing down after about 1h45m of runtime. At this point I stopped the `divisorsearch` container and ran the timing test container. And indeed, the upsert step has a strangely slow pattern (the numbers are in seconds):

```Running sample 0
55.63926291465759
Running sample 1
61.28182792663574
Running sample 2
245.36470413208008  # 4 minutes
Running sample 3
1683.663686990738   # 28 minutes
```

At this point I killed the timing test. The initial few writes, while not great, are not as concerning as the massive degradation in performance over four writes. At this point I’m convinced that the database is the problem. So I read a lot about Postgres performance tuning (e.g. these docs and this blog post), which I had never done before.

It appears that the way Postgres inserts work is by writing to a “write cache,” and then later a background process copies the content of the write cache to disk. This makes sense because disk writes are slow. But if the write cache is full, then Postgres will wait until the write cache gets emptied before it can continue. This seems likely here: the first few writes are fast, and then when the cache fills up later writes take eons.

I also learned that the `PRIMARY KEY` part of our Postgres schema incurs a penalty to write performance, because in order to enforce the uniqueness Postgres needs to maintain a data structure and update it on every write. I can imagine that, because an `mpz` (unbounded integer) is our primary key, that data structure might be unprepared to handle such large keys. This might explain why these writes are taking so long in the first place, given that each write itself is relatively small (~2 MiB). In our case we can probably get by without a primary key on this table, and instead put any uniqueness constraints needed on the search metadata table.

So let’s change one of these and see if it fixes the problem. First we can drop the primary key constraint, which we can do in situ by running the query

```alter table riemanndivisorsums drop constraint divisor_sum_pk;
```

This query takes a long time, which seems like a good sign. Perhaps it’s deleting a lot of data. Once that’s done, I removed the `ON CONFLICT` clause from the upsert command (it becomes just an insert command), and then rebuild and re-run the timing container. This shows that each insert takes only 3-4 seconds, which is much more reasonable.

```Running sample 0
3.8325071334838867
Running sample 1
3.7897982597351074
Running sample 2
3.7978150844573975
Running sample 3
3.810023784637451
Running sample 4
3.8057897090911865
```

I am still curious about the possible fix by increasing the cache size, but my guess is that without the primary key change, increasing the cache size would just delay the application from slowing instead of fixing the problem permanently, so for now I will not touch it. Perhaps a reader with more experience with Postgres tuning would know better.

## Updating the application

So with this new understanding, how should we update the application? Will anything go wrong if we delete the primary key on the `RiemannDivisorSums` table?

As suggested by our need to remove the `ON CONFLICT` clause, we might end up with duplicate rows. That is low risk, but scarier is if we stop the search mid-block and restart it, then if we’re really unlucky, we might get into a state where we skip a block and don’t notice, and that block contains the unique counterexample to the Riemann Hypothesis! We simply can’t let that happen. This is far too important. To be fair, this is a risk even before we remove the primary key. I’m just realizing it now.

We’ll rearchitect the system to mitigate that in a future post, and it will double as providing a means to scale horizontally. For now, this pull request removes the primary key constraint and does a bit of superficial cleanup. Now let’s re-run it!

## Out of RAM again

This time, it ran for three hours before the divisorsearch container hit the 15 GiB RAM limit and crashed. Restarting the container and watching `docker stats` showed that it plain ran out of RAM. This is a much easier problem to solve because it involves only one container, and the virtual machine doesn’t slow to a crawl when it occurs, the container just stops running.

A quick python memory profile reveals the top 10 memory usages by line, the top being `search_strategy.py:119` clocking in at about 2.2 GiB.

```[ Top 10 ]
riemann/search_strategy.py:119: size=2245 MiB, count=9394253, average=251 B
venv/lib/python3.7/site-packages/llvmlite/ir/values.py:224: size=29.6 MiB, count=9224, average=3367 B
<string>:2: size=26.7 MiB, count=499846, average=56 B
riemann/superabundant.py:67: size=19.1 MiB, count=500003, average=40 B
riemann/superabundant.py:69: size=15.3 MiB, count=250000, average=64 B
riemann/superabundant.py:70: size=13.4 MiB, count=250000, average=56 B
venv/lib/python3.7/site-packages/llvmlite/ir/_utils.py:48: size=3491 KiB, count=6411, average=558 B
venv/lib/python3.7/site-packages/llvmlite/ir/values.py:215: size=2150 KiB, count=19267, average=114 B
riemann/search_strategy.py:109: size=2066 KiB, count=1, average=2066 KiB
venv/lib/python3.7/site-packages/llvmlite/ir/_utils.py:58: size=1135 KiB, count=2018, average=576 B
```

This line, not surprisingly, is the computation of the full list of partitions of $n$. The length of this list grows superpolynomially in $n$, which explains why it takes up all the RAM.

Since we only need to look at at most `batch_size` different partitions of $n$ at a given time, we can probably fix this by adding a layer of indirection so that new sections of the partition list are generated on the fly and old sections are forgotten once the search strategy is done with them.

This pull request does that, and running the memory test again shows the 2.5 GiB line above reduced to 0.5 GiB. The pull request description also has some thoughts about the compute/memory tradeoff being made by this choice, and thoughts about how to improve the compute side.

So let’s run it again!! This time it runs for about 19 hours before crashing, as shown by the EC2 CPU usage metrics.

According to the `SearchMetadata` table, we got up to part way through $n=87$

```divisor=# select * from searchmetadata order by start_time desc limit 3;
start_time         |          end_time          |       search_state_type       | starting_search_state | ending_search_state
----------------------------+----------------------------+-------------------------------+-----------------------+---------------------
2021-01-27 00:00:11.59719  | 2021-01-27 00:01:14.709715 | SuperabundantEnumerationIndex | 87,15203332           | 87,15453332
2021-01-26 23:59:01.285508 | 2021-01-27 00:00:11.596075 | SuperabundantEnumerationIndex | 87,14953332           | 87,15203332
2021-01-26 23:57:59.809282 | 2021-01-26 23:59:01.284257 | SuperabundantEnumerationIndex | 87,14703332           | 87,14953332
```

Logging into the server and running `df -h` we can see the reason for the crash is that the disk is full. We filled up a 60 GiB SSD full of Riemann divisor sums. It’s quite an achievement. I’d like to thank my producers, the support staff, my mom, and the Academy.

But seriously, this is the best case scenario. Our limiting factor now is just how much we want to pay for disk space (and/or, any tricks we come up with to reduce disk space).

Analyzing the results, it seems we currently have 949 examples of numbers with witness values bigger than 1.767. Here are the top few:

```                                                                                       n                                                                                        |                                                                                   divisor_sum                                                                                   |   witness_value
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+--------------------
533187564151227457465199401229454876347036513892234205802944360099435118364718466037392872608220305945979716166395732328054742493039981726997486787797703088097204529280000    | 5634790045188254963919691913741193557789227457256740288527114259589826494137632464425981545755572664137227875215269764405574488353138596297944567958732800000000000000000000    | 1.7689901658397056
1392134632248635659178066321747923959130643639405311242305337754828812319490126543178571468950966856255821713228187290673772173611070448373361584302343872292681996160000      | 14673932409344413968540864358701024890076113169939427834706026717681839828483417876109326942071803812857364258373098344806183563419631761192563979059200000000000000000000      | 1.7688977455109272
3673178449204843427910465228886342900080853929829317262019360830682882109472629401526573796704398037614305311947723722094385682351109362462695473093255599716839040000         | 38615611603537931496160169365002697079147666236682704828173754520215367969693204937129807742294220560150958574666048275805746219525346739980431523840000000000000000000         | 1.7688303488154073
2784269264497271318356132643495847918261287278810622484610675509657624638980253086357142937901933712511643426456374581347544347222140896746723168604687744585363992320000      | 29355033324995298095346770663840105972086801871471400577978104254473441181064775868425839681379597759477726740614478613571725301516068423098949435392000000000000000000000      |  1.768798862584946
9847663402693950208875241900499578820592101688550448423644399009873678577674609655567221975078815114247467324256631962719532660458738237165403413118647720420480000            | 103250298405181635016471041082894911976330658386852151946988648449773711148912312666122480594369573690243204745096385764186487217982210534707036160000000000000000000           | 1.7687597437473672
```

The best so far achieves 1.76899. Recall out last best number achieved a witness value of 1.7679, a modest improvement if we hope to get to 1.82 (or even the supposedly infinitely many examples with witness value bigger than 1.81!).

## What’s next?

We’re not stopping until we disprove the Riemann hypothesis. So strap in, this is going to be a long series.

Since the limiting factor in our search is currently storage space, I’d like to spend some time coming up with a means to avoid storing the witness value and divisor sum for every single number in our search strategy, but still maintain the ability to claim our result (no counterexamples below N) while providing the means for a skeptical person to verify our claims as much as they desire without also storing the entire search space on disk.

Once we free ourselves from disk space constraints, the second limiting factor is that our search uses only a single CPU at a time. We should rearchitect the system so that we can scale horizontally, meaning we can increase the search rate by using more computers. We can do this by turning our system into a “worker” model, in which there is a continually growing queue of search blocks to process, and each worker machine asks for the next block to compute, computes it, and then stores the result in the database. There are some tricks required to ensure the workers don’t step on each other’s toes, but it’s a pretty standard computational model.

There’s also the possibility of analyzing our existing database to try to identify what properties of the top $n$ values make their witness values so large. We could do this by re-computing the prime factorization of each of the numbers and squinting at it very hard. The benefit of doing this would ultimately be to design a new `SearchStrategy` that might find larger witness values faster than the superabundant enumeration. That approach also has the risk that, without a proof that the new search strategy is exhaustive, we might accidentally skip over the unique counterexample to the Riemann hypothesis!

And then, of course, there’s the business of a front-end and plotting. But I’m having a bit too much fun with the back end stuff to worry about that for now. Protest in the comments if you prefer I work on visualization instead.

## Postscript: a false start

Here’s one thing I tried during debugging that turned out to be a dead end: the linux `perf` tool. It is way more detailed than I can understand, but it provides a means to let me know if the program is stuck in the OS kernel or whether it’s the applications that are slow. I ran the perf tool when the deployed application was in the “good performance” regime (early in its run). I saw something like this:

```Samples: 59K of event 'cpu-clock:pppH', Event count (approx.): 119984000000
49.13%  swapper          [kernel.kallsyms]                            [k] native_safe_halt
8.04%  python3          libpython3.7m.so.1.0                         [.] _PyEval_EvalFrameDefault
0.86%  python3          libc-2.28.so                                 [.] realloc
0.86%  python3          libpython3.7m.so.1.0                         [.] 0x0000000000139736
0.77%  python3          libpython3.7m.so.1.0                         [.] PyTuple_New
0.74%  python3          libpython3.7m.so.1.0                         [.] PyType_IsSubtype

```

It says half the time is spent by something called “swapper” doing something called `native_safe_halt`, and this in kernel mode (`[k]`), i.e., run by the OS. The rest of the list is dominated by python3 and postgres doing stuff like `_PyDict_LoadGlobal` which I assume is useful Python work. When I look up `native_safe_halt` (finding an explanation is surprisingly hard), I learn that this indicates the system is doing nothing. So 49% of the time the CPU is idle. This fits with good behavior in our case, because each docker container gets one of the two CPUs on the host, and the postgres container is most often doing nothing while waiting for the search container to send it data. This also matches the CPU utilization on the EC2 dashboard. So everything appears in order.

I ran perf again after the application started slowing down, and I saw that `native_safe_halt` is at 95%. This tells me nothing new, sadly. I also tried running it during the timing test and saw about 40% usage of some symbols like `__softirqentry_text_start` and `__lock_text_start`. Google failed to help me understand what that was, so it was a dead end.

# Searching for RH Counterexamples — Deploying with Docker

We’re ironically searching for counterexamples to the Riemann Hypothesis.

In this article we’ll deploy the application on a server, so that it can search for RH counterexamples even when I close my laptop.

## Servers and containers

When deploying applications to servers, reproducibility is crucial. You don’t want your application to depend on the details of the computer it’s running on. This is a higher-level version of the same principle behind Python virtual environments, but it applies to collections of programs, possibly written in different languages and running on different computers. In our case, we have a postgres database, the pgmp extension, the `populate_database` program, and plans for a web server.

The principle of an application not depending on the system it’s running on is called hermeticity (the noun form of hermetic, meaning air-tight). Hermeticity is good for the following reasons. When a server crashes, you don’t have to remember what you did to set it up. Instead you run a build/install that works on any machine. Newcomers also don’t have to guess what unknown aspects of a running server are sensitive. Another benefit is that you can test it on your local machine identically to how it will run in production. It also allows you to easily migrate from one cloud provider to another, which allows you to defer expensive commitments until you have more information about your application’s needs. Finally, if you have multiple applications running on the same server, you don’t want to have their needs conflict with each other, which can happen easily if two applications have dependencies that transitively depend on different versions of the same software. This is called “dependency hell.” In all of these, you protect yourself from becoming dependent on arbitrary choices you made before you knew better.

One industry-strength approach to hermeticity is to use containers. A container is a virtual machine devoted to running a single program, with explicitly-defined exposure to the outside world. We will set up three containers: one to run the database, one for the search application, and (later) one for a web server. We’ll start by deploying them all on the same machine, but could also deploy them to different machines. Docker is a popular containerization system. Before going on, I should stress that Docker, while I like it, is not sacred by any means. In a decade Docker may disappear, but the principle of hermeticity and the need for reproducible deployments will persist.

Docker allows you to describe your container by first starting from an existing (trusted) container, such as one that has an operating system and postgres already installed, and extend it for your application. This includes installing dependencies, fetching the application code from git, copying files into the container from the host system, exposing the container’s network ports, and launching the application. You save the commands that accomplish that in a Dockerfile with some special syntax. To deploy it, you copy the Dockerfile to the server (say, via git) and run docker commands to launch the container. You only have to get the Dockerfile right once, you can test it locally, and then it will work on any server just the same. The only caveat I’ve seen here is that if you migrate to a server with a different processor architecture, the install script (in our case, `pip install numba`) may fail to find a pre-compiled binary for the target architecture, and it may fall back to compiling from source, which can add additional requirements or force you to change which OS your container is derived from.

This reduces our “set up a new server” script to just a few operations: (1) install docker (2) fetch the repository (3) launch the docker containers from their respective Dockerfiles. In my experience, writing a Dockerfile is no small task, but figuring out how to install stuff is awful in all cases, and doing it for Docker gives you an artifact tracing the steps, and a reasonable expectation of not having to do it again.

Thankfully, you dear readers can skip my head-banging and see the Dockerfiles after I figured it out.

## The Postgres Dockerfile

This commit adds a Dockerfile for the database, and makes some small changes to the project to allow it to run. It has only 15 lines, but it took me a few hours to figure out. The process was similar to installing confusing software on your own machine: try to install, see some error like `"missing postgres.h`“, go hunt around on the internet to figure out what you have to install to get past the error, and repeat.

Let’s go through each line of the Dockerfile.

```FROM postgres:12
```

The first line defines the container image that this container starts from, which is officially maintained by the Postgres team. Looking at their Dockerfile, it starts from `debian:buster-slim`, which is a Debian Linux instance that is “slimmed” down to be suitable for docker containers, meaning it has few packages pre-installed. Most importantly, “Debian” tells us what package manager to use (apt-get) in our Dockerfile.

It’s also worth noting at this point that, when docker builds the container, each command in a docker file results in a new image. An image is a serialized copy of all the data in a docker container, so that it can be started or extended easily. And if you change a line halfway through your Dockerfile, docker only has to rebuild images from that step onward. You can publish images on the web, and other docker users can use them as a base. This is like forking a project on Github, and is exactly what happens when Docker executes `FROM postgres:12`.

```ENV POSTGRES_USER docker
ENV POSTGRES_DB divisor
```

These lines declare configuration for the database that the base postgres image will create when the container is started. The variable names are described in the “Environment Variables” section of the Postgres image’s documentation. The `ENV` command tells docker to instantiate environment variables (like the `PATH` variable in a terminal shell), that running programs can access. I’m insecurely showing the password and username here because the server the docker containers will run on won’t yet expose anything to the outside world. Later in this post you will see how to pass an environment variable from the docker command line when the container is run, and you would use something close to that to set configuration secrets securely.

```RUN apt-get update \
&& apt-get install -y pgxnclient build-essential libgmp3-dev postgresql-server-dev-12 libmpc-dev
```

The `RUN` command allows you to run any shell command you’d like, in this case a command to update apt and install the dependencies needed to build the pgmp extension. This includes `gcc` and `make` via `build-essential`, and the gmp-specific libraries.

```RUN apt-get install -y python3.7 python3-setuptools python3-pip python-pip python3.7-dev \
&& pip3 install wheel \
&& pip install six
```

Next we do something a bit strange. We install python3.7 and pip (because we will need to pip3 install our project’s requirements.txt), but also python2’s pip. Here’s what’s going on. The pgmp postgres extension needs to be built from source, and it has a dependency on python2.7 and the python2-six library. So the first RUN line here installs all the python-related tools we need.

```RUN pgxn install pgmp
```

Then we install the pgmp extension.

```COPY . /divisor
WORKDIR "/divisor"
```

These next two lines copy the current directory on the host machine to the container’s file system, and sets the working directory for all future commands to that directory. Note that whenever the contents of our project change, docker needs to rebuild the image from this step because any subsequent steps like `pip install -r requirements.txt` might have a different outcome.

```RUN python3 -m pip install --upgrade pip
RUN pip3 install -r requirements.txt
```

Next we upgrade pip (which is oddly required for the numba dependency, though I can’t re-find the Github issue where I discovered this) and install the python dependencies for the project. The only reason this is required is because we included the database schema setup in the python script `riemann/postgres_batabase.py`. So this makes the container a bit more complicated than absolutely necessary. It can be improved later if need be.

```ENV PGUSER=docker
ENV PGDATABASE=divisor
```

These next lines are environment variables used by the `psycopg2` python library to infer how to connect to postgres if no database spec is passed in. It would be nice if this was shared with the postgres environment variables, but duplicating it is no problem.

```COPY setup_schema.sh /docker-entrypoint-initdb.d/
```

The last line copies a script to a special directory specified by the base postgres Dockerfile. The base dockerfile specifies that any scripts in this directory will be run when the container is started up. In our case, we just call the (idempotent) command to create the database. In a normal container we might specify a command to run when the container is started (our search container, defined next, will do this), but the postgres base image handles this for us by starting the postgres database and exposing the right ports.

Finally we can build and run the container

```docker build -t divisordb -f divisordb.Dockerfile .
#  ... lots of output ...

docker run -d -p 5432:5432 --name divisordb divisordb:latest
```

After the `docker build` command—which will take a while—you will be able to see the built images by running `docker images`, and the final image will have a special tag `divisordb`. The run command additionally tells docker to run the container as a daemon (a.k.a. in the background) with `-d` and to `-p` to publish port 5432 on the host machine and map it to 5432 on the container. This allows external programs and programs on other computers to talk to the container by hitting `0.0.0.0:5432`. It also allows other containers to talk to this container, but as we’ll see shortly that requires a bit more work, because inside a container `0.0.0.0` means the container, not the host machine.

Finally, one can run the following code on the host machine to check that the database is accepting connections.

```pg_isready --host 0.0.0.0 --username docker --port 5432 --dbname divisor
```

If you want to get into the database to run queries, you can run `psql` with the same flags as `pg_isready`, or manually enter the container with `docker exec -it divisordb bash` and run `psql` from there.

```psql --host 0.0.0.0 --username docker --port 5432 --dbname divisor
divisor=# \d
List of relations
Schema |        Name        | Type  | Owner
--------+--------------------+-------+--------
public | riemanndivisorsums | table | docker
public | searchmetadata     | table | docker
(2 rows)
```

Look at that. You wanted to disprove the Riemann Hypothesis, and here you are running docker containers.

## The Search Container

Next we’ll add a container for the main search application. Before we do this, it will help to make the main entry point to the program a little bit simpler. This commit modifies `populate_database.py`‘s main routine to use argparse and some sensible defaults. Now we can run the application with just `python -m riemann.populate_database`.

Then the Dockerfile for the search part is defined in this commit. I’ve copied it below. It’s much simpler than the database, but somehow took just as long for me to build as the database Dockerfile, because I originally chose a base image called “alpine” that is (unknown to me at the time) really bad for Python if your dependencies have compiled C code, like numba does.

```FROM python:3.7-slim-buster

RUN apt-get update \
&& apt-get install -y build-essential libgmp3-dev libmpc-dev

COPY . /divisor
WORKDIR "/divisor"

RUN pip3 install -r requirements.txt

ENV PGUSER=docker
ENV PGDATABASE=divisor

ENTRYPOINT ["python3", "-m", "riemann.populate_database"]
```

The base image is again Debian, with Python3.7 pre-installed.

Then we can build it and (almost) run it

```docker build -t divisorsearch -f divisorsearch.Dockerfile .
docker run -d --name divisorsearch --env PGHOST="\$PGHOST" divisorsearch:latest
```

What’s missing here is the `PGHOST` environment variable, which `psycopg2` uses to find the database. The problem is, inside the container “localhost” and `0.0.0.0` are interpreted by the operating system to mean the container itself, not the host machine. To get around this problem, docker maintains IP addresses for each docker container, and uses those to route network requests between containers. The `docker inspect` command exposes information about this. Here’s a sample of the output

```\$ docker inspect divisordb
[
{
"Id": "f731a78bde50be3de1d77ae1cff6d23c7fe21d4dbe6a82b31332c3ef3f6bbbb4",
"Path": "docker-entrypoint.sh",
"Args": [
"postgres"
],
"State": {
"Status": "running",
"Running": true,
"Paused": false,
...
},
...
"NetworkSettings": {
...
"Ports": {
"5432/tcp": [
{
"HostIp": "0.0.0.0",
"HostPort": "5432"
}
]
},
...
...
}
}
]
```

The part that matters for us is the ip address, and the following extracts it to the environment variable `PGHOST`.

```export PGHOST=\$(docker inspect -f "{{ .NetworkSettings.IPAddress }}" divisordb)
```

Once the two containers are running—see `docker ps` for the running containers, `docker ps -a` to see any containers that were killed due to an error, and `docker logs` to see the container’s logged output—you can check the database to see it’s being populated.

```divisor=# select * from SearchMetadata order by start_time desc limit 10;
start_time         |          end_time          |       search_state_type       | starting_search_state | ending_search_state
----------------------------+----------------------------+-------------------------------+-----------------------+---------------------
2020-12-27 03:10:01.256996 | 2020-12-27 03:10:03.594773 | SuperabundantEnumerationIndex | 29,1541               | 31,1372
2020-12-27 03:09:59.160157 | 2020-12-27 03:10:01.253247 | SuperabundantEnumerationIndex | 26,705                | 29,1541
2020-12-27 03:09:52.035991 | 2020-12-27 03:09:59.156464 | SuperabundantEnumerationIndex | 1,0                   | 26,705
```

## Ship it!

I have an AWS account, so let’s use Amazon for this. Rather than try the newfangled beanstalks or lightsails or whatever AWS-specific frameworks they’re trying to sell, for now I’ll provision a single Ubuntu EC2 server and run everything on it. I picked a t2.micro for testing (which is free). There’s a bit of setup to configure and launch the server—such as picking the server image, downloading an ssh key, and finding the IP address. I’ll skip those details since they are not (yet) relevant to the engineering process.

Once I have my server, I can ssh in, install docker, git clone the project, and run the deploy script.

```# install docker, see get.docker.com
curl -fsSL https://get.docker.com -o get-docker.sh
sudo sh get-docker.sh
sudo usermod -aG docker ubuntu

# log out and log back in

git clone https://github.com/j2kun/riemann-divisor-sum && cd riemann-divisor-sum
bash deploy.sh
```

And it works!

Sadly, within an hour the `divisorsearch` container crashes because the instance runs out of RAM and CPU. Upgrading to a t2.medium (4 GiB RAM), it goes for about 2 hours before exhausting RAM. We could profile it and find the memory hotspots, but instead let’s apply a theorem due to billionaire mathematician Jim Simons: throw money at the problem. Upgrading to a r5.large (16 GiB RAM), and it runs comfortably all day.

Four days later, logging back into the VM and and I notice things are sluggish, even though the docker instance isn’t exhausting the total available RAM or CPU. `docker stats` also shows low CPU usage on `divisorsearch`. The database shows that it has only got up to 75 divisors, which is just as far as it got when I ran it (not in Docker) on my laptop for a few hours in the last article.

Something is amiss, and we’ll explore what happened next time.

## Notes

A few notes on improvements that didn’t make it into this article.

In our deployment, we rebuild the docker containers each time, even when nothing changes. What one could do instead is store the built images in what’s called a container registry, and pull them instead of re-building them on every deploy. This would only save us a few minutes of waiting, but is generally good practice.

We could also use docker compose and a corresponding configuration file to coordinate launching a collection of containers that have dependencies on each other. For our case, the `divisorsearch` container depended on the `divisordb` container, and our startup script added a `sleep 5` to ensure the latter was running before starting the former. `docker compose` would automatically handle that, as well as the configuration for naming, resource limits, etc. With only two containers it’s not that much more convenient, given that `docker compose` is an extra layer of indirection to learn that hides the lower-level commands.

In this article we deployed a single database container and a single “search” container. Most of the time the database container is sitting idle while the search container does its magic. If we wanted to scale up, an obvious way would be to have multiple workers. But it would require some decent feature work. A sketch: reorganize the `SearchMetadata` table so that it contains a state attribute, like “not started”, “started”, or “finished,” then add functionality so that a worker (atomically) asks for the oldest “not started” block and updates the row’s state to “started.” When a worker finishes a block, it updates the database and marks the block as finished. If no “not started” blocks are found, the worker proceeds to create some number of new “not started” blocks. There are details to be ironed out around race conditions between multiple workers, but Postgres is designed to make such things straightforward.

Finally, we could reduce the database size by keeping track of a summary of a search block instead of storing all the data in the block. For example, we could record the `n` and `witness_value` corresponding to the largest `witness_value` in a block, instead of saving every `n` and every `witness_value`. In order for this to be usable—i.e., for us to be able to say “we checked all possible $n < M$ and found no counterexamples”—we’d want to provide a means to verify the approach, say, by randomly verifying the claimed maxima of a random subset of blocks. However, doing this also precludes the ability to analyze patterns that look at all the data. So it’s a tradeoff.

# Searching for RH Counterexamples — Unbounded Integers

We’re ironically searching for counterexamples to the Riemann Hypothesis.

In the last article, we improved our naive search from “try all positive integers” to enumerate a subset of integers (superabundant numbers), which RH counterexamples are guaranteed to be among. These numbers grow large, fast, and we quickly reached the limit of what 64 bit integers can store.

Unbounded integer arithmetic is possible on computers, but it requires a special software implementation. In brief, you represent numbers in base-N for some large N (say, $2^{32}$), and then use a 32-bit integer for each digit. Arithmetic on such quantities emulates a ripple-carry adder, which naturally requires linear time in the number of digits of each operand. Artem Golubin has a nice explanation of how Python does it internally.

So Python can handle unbounded integer arithmetic, but neither numba nor our database engine do. Those both crash when exceeding 64-bit integers This is a problem because we won’t be able to store the results of our search without being able to put it in a database. This leaves us with a classic software engineering problem. What’s the path forward?

## Exploring Alternatives

The impulse answer is to do as little as possible to make the damn thing work. In a situation where the software you’re writing is a prototype, and you expect it to be rewritten from scratch in the future, this is an acceptable attitude. That said, experienced engineers would caution you that, all too often, such “prototypes” are copy-pasted to become janky mission-critical systems for years.

In pretending this is the “real thing,” let’s do what real engineers would do and scope out some alternatives before diving in. The two aspects are our database and the use of numba for performance.

Let’s start with the database. A quick and dirty option: store all numbers as text strings in the database. There’s no limit on the size of the number in that case. The benefit: we don’t need to use a different database engine, and most of our code stays the same. The cost: we can’t use numeric operations in database queries, which would make further analysis and fetching awkward. In particular, we can’t even apply sorting operations, since text strings are sorted lexicographically (e.g., 100, 25) while numbers are sorted by magnitude (25, 100). Note, we applied this “numbers as text” idea to the problem of serializing the search state, and it was hacky there, too.

A second option is to find a database engine with direct support for unbounded-integer arithmetic. The benefit: fast database queries and the confidence that it will support future use cases well. The cost: if our existing sqlite-based interface doesn’t work with the new database engine, we’d have to write another implementation of our database interface.

For numba, we have at least three options. First, fall back to native python arithmetic, which is slow. Second, implement arbitrary-precision arithmetic in Python in a way that numba can compile it. Third, find (or implement) a C-implementation of arbitrary precision integer arithmetic, provide Python bindings, and optionally see if it can work with (or replace) numba. As I write this I haven’t yet tried any of these options. My intuition tells me the best way to go would be to find “proper” support for arbitrary precision integers.

For the database, I recall that the Postgres database engine supports various extensions, for example this extension that adds support for geographic objects. Postgres’s extension framework demonstrates an important software engineering principle that many of the best projects follow: “closed for modification, open for extension.” That is, Postgres is designed so that others can contribute new features to Postgres without requiring the Postgres team to do anything special—specifically, they don’t have to change Postgres to accommodate it. The name for this sometimes goes by extensions, or plug-ins, hooks, or (at a lower level) callbacks. Github Actions is a good example of this.

Geographic objects are almost certainly more complicated than arbitrary precision integers, so chances are good a Postgres extension exists for the latter. Incorporating it would involve migrating to Postgres, finding and installing that extension, and then converting the C library representation above to whatever representation Postgres accepts in a query.

A good route will also ensure that we need not change our tests too much, since all we’re doing here is modifying implementations. We’ll see how well that holds up.

## gmp and pgmp

After some digging, I found GMP (GNU Multiple Precision), a C library written by Torbjörn Granlund. It has a Python bindings library called gmpy that allows Python to use an “mpz” (“Multiple Precision $\mathbb{Z}$“) type as a drop-in replacement for Python integers. And I found a PostgreSQL extension called pgmp. The standard Python library for Postgres is psycopg2, which was written by the same person who wrote pgmp, Daniele Varrazzo.

To start, I ran a timing test of gmpy, which proves to be as fast as numba. This pull request has the details.

It took a small bit of kicking to get pgmp to install, but then I made a test database that uses the new column type mpz and stores the value $2^{513}$.

```postgres=# create database pgmp_test;
CREATE DATABASE
postgres=# \connect pgmp_test;
You are now connected to database "pgmp_test" as user "jeremy".
pgmp_test=# CREATE EXTENSION pgmp;
CREATE EXTENSION
pgmp_test=# create table test_table (id int4, value mpz);
CREATE TABLE
pgmp_test=# insert into test_table
pgmp_test-# values (1, 2::mpz ^ 513);
INSERT 0 1
pgmp_test=# select * from test_table;
id |                                                                            value
----+-------------------------------------------------------------------------------------------------------------------------------------------------------------
1 | 26815615859885194199148049996411692254958731641184786755447122887443528060147093953603748596333806855380063716372972101707507765623893139892867298012168192
(1 row)
```

Now I’m pretty confident this approach will work.

This pull request includes the necessary commits to add a postgres implementation of our database interface, add tests (which is a minor nuisance).

Then this pull request converts the main divisor computation functions to use gmpy, and this final commit converts the main program to use the postgres database.

This exposed one bug, that I wasn’t converting the new mpz types properly in the postgres sql query. This commit fixes it, and this commit adds a regression test to catch that specific error going forward.

## Results and next steps

With all that work, I ran the counterexample search for a few hours.

When I stopped it, it had checked all possibly-superabundant numbers whose prime factorizations have at most 75 prime factors, including multiplicity. Since all possible counterexamples to the RH must be superabundant, and all superabundant numbers have the aforementioned special prime factorization, we can say it more simply. I ruled out all positive integers whose prime factorization has at most 75 factors.

The top 10 are:

```divisor=# select n, witness_value
from RiemannDivisorSums
where witness_value > 1.7 and n > 5040
order by witness_value desc
limit 10;
n                                                                          |   witness_value
----------------------------------------------------------------------------------------------------------------------------------------------------+--------------------
7837096340441581730115353880089927210115664131849557062713735873563599357074419807246597145310377220030504976899588686851652680862494751024960000  | 1.7679071291526642
49445402778811241199465955079431717413978953513246416799455746836363402883750282695562127099750014006501608687063651021146073696293342277760000    |  1.767864530684858
24722701389405620599732977539715858706989476756623208399727873418181701441875141347781063549875007003250804343531825510573036848146671138880000    |  1.767645098171234
157972532839652527793820942745788234549453525601426251755449670403716942120607931934703281468849885004797471843653837128262216282087355520000      | 1.7676163327497005
2149800120817880052150693699105726844086041457097670295628510732015800125380447073720092482597826695934852551611463087875916247664927925120000     |  1.767592584103948
340743319149633988265884951308257704787637570949980741857118951024504319872800861184634658491755531305674129430416899428332725254891076131520000   |  1.767582883432923
23511289021324745190346061640269781630346992395548671188141207620690798071223259421739791435931131660091514930698766060554958042587484253074880000 | 1.7674462177172812
507950266365442211555694349664913937458049921547994378634886400011951582381375986928306371282475514484879330686989829994412271003496320000         | 1.7674395010995763
78986266419826263896910471372894117274726762800713125877724835201858471060303965967351640734424942502398735921826918564131108141043677760000       | 1.7674104158678667
6868370993028370773644388815034271067367544591366358771976072626248562700895997040639273107341299348034672688854514657750531142699450240000        | 1.7674059308384011
```

This is new. We’ve found quite a few numbers that have a better witness value than $n = 10080$ which achieves ~1.7558. The best is

78370963404415817301153538800899272101156641318495
57062713735873563599357074419807246597145310377220
030504976899588686851652680862494751024960000

which achieves ~1.7679. Recall the 1.781 threshold needed to be a RH counterexample. We’re about 50% of the way toward disproving RH. How much more work could it take?

But seriously, what’s next with this project? For one, even though we have some monstrous numbers and their divisor sums and witness values, it’s hard to see the patterns in them through a SQL queries. It would be nicer to make some pretty plots.

I could also take a step back and see what could be improved from a software engineering standpoint. For one, not all parts of the application are tested, and tests aren’t automatically run when I make changes. This enabled the bug above where I didn’t properly convert mpz types before passing them to SQL upsert statements. For two, while I have been using type annotations in some places, they aren’t checked, and the switch to mpz has almost certainly made many of the type hints incorrect. I could fix that and set up a test that type checks.

Finally, in the interest of completeness, I could set up a front end that displays some summary of the data, and then deploy the whole application so that it has a continuously-running background search, along with a website that shows how far along the search is. Based on how long the SQL statement to find the top 10 witness values took, this would also likely require some caching, which fits snugly in the class of typical software engineering problems.

Let me know what you’re interested in.