（评论）

（评论）
(comments)

原始链接: https://news.ycombinator.com/item?id=39955944

由于对视觉设计器和布局引擎感到不满，该用户更喜欢使用简单的 Markdown 表而不是图表来创建数据模型和实体关系图 (ERD)。他们重视表格模式和关系而不是图表美观。他们利用代码块和自行创建的速记法，以文本方式记录字段、类型和属性。为其他人生成图像是在设计迭代后完成的。尽管 D2、Mermaid、PlantUML 等当前工具面临挑战，但他们的目标是创建一个基于文本的、tree-sitter diffable 和 Git 存储的系统，用于创建和比较架构图。他们设想了一个机器学习模型，要求图表的相关部分，生成不同的输出，并应用各种样式。此外，他们还希望能够在版本和行业标准之间进行比较。这可能包括发出 SQL 和 TypeScript/C# 类，以便于实现和分析。他们建议改进描述，以实现更清晰的沟通，并承认标签分配方面的混乱。他们欣赏 SchemaSpy 等工具过去的推荐，但更喜欢专门针对关系数据库而不是 Java 生成文档。总的来说，他们寻求一种本地的、不依赖于云的工具来创建、保存和操作数据库架构图。

Having spent fair amount of time on inf.architectural position, I’d really love these tools to also adapt some sort of markup. Sequencediagram.org being perfect example (and others alike) where it takes seconds to put a diagram that otherwise takes hours to prepare. Sadly I know of no standard and lightweight markup for E/R modelling…

Mermaid is good for flowcharts, but we still lack reasonably good one for BPMN and enterprise diagrams. (BPMN.io is great but lacks its own lang)

I have found that diagrams tend not to be worth their effort for data models and ERDs.

I spend far too much time either:

- working in a visual designer that slows me down

Or

- fighting the layout engine of a markup language and image compiler with lots of directives to try to get the diagram to lay out anywhere close to reasonably.

The most valuable information, to me, is the table schema and the collection of relationships. Toward this end, I use a markdown table of three columns, (Field, Type, Attributes) and a code block with my own shorthand per below:

    Table[field] -1:N- Table2[field2]
    Detail[HeaderId] -N:1- Header[Id]

All text, no generated pictures. If I need to generate from DDL, that's pretty straightforward. But I can also mock up new designs at typing speed.

Diffs are incredibly readable.

If I need to make pictures for someone else, I iterate design in a markdown file and then once it's good I will spend the time required to make a picture based off the text.

I must admit, I start to question why the pictures are necessary.

> I start to question why the pictures are necessary

I think I agree. My habit is to keep DDL in version control, one file per database object. If I'm interested in a table, column, or its relationships, I can grep for names, look at the files and see the datatypes, foreign keys, etc.

I can also by grepping find any stored procedure or function that references a table or column because the code for those is in version control also.

I've tried diagrams and while they are convenient at a fairly high level they can't really show you everything and become unweildy if they try. I've found them of such limited benefit that I just work with text/code. I might make a simple diagram if I have to present/talk about a change.

Yeah, the developer experience (DevEx) around diagrams is still really poor, despite some sterling efforts by folks behind Mermaid, PlantUML, D2, etc etc.

I want something that will use (text based, tree-sitter diffable, and git stored) architecture diagrams as its source of truth. Then compare those diagrams with actual code, and give us meaningful/useful divergence metrics. And from the source of truth, apply style transforms to display in different ways. D2 can do something like that last bit, with its different layout engines.

It feels like we already have the information available: we have (terraform) code that produces infrastructure, and (nodejs) code that produces applications. So how can take that structural information and turn it into something useful? I kind of like the idea of a ML model that gives you some standard views of your diagrams, then asks you which bits are useful, and which bits you want to hide/discard from the final output. Going back and forth with that a bit would produce a JSON file that tells the transform engine what to display. I don't know if an LLM can be prompted to do something like that, but a model-assisted workflow might be useful.

I also want to be able to compare before and after architectural decisions. Or compare the infrastructure and application patterns against industry best practices. Plus being able to automatically spit out updated, accurate diagrams for leadership would be great.

Are there any advantages that could come from encouraging a tool like this to emit both SQL and strongly-typed classes representing the data in say TypeScript and C#?

Fantastic. I was literally just looking for something like this. I'm using it now.

One thing that may be me getting my brain reversed, but when you set the "cardinality" of a relationship the "1" and "n" markings on the connecting line seem to be reversed?

Thanks for the feedback! The "1" and "n" labels are assigned first to the foreign key then to the primary key, e.g. if you pick "Many to one" the FK side will be "n" and the PK will be "1". Let me know if you find this unintuitive or if I didn't get the question right.

This looks great! However, a pet peeve of mine that I always share with my students is that ER diagrams do not have foreign keys. ER diagrams represent relationships with a diamond connected by lines to the related entities. It looks like what this is producing is not really an ER diagram.

The big diamonds are specifically the Chen notation of ER diagrams. Which is popular in academia because that's how Professor Chen first described ER models, but seems pretty unpopular everywhere else because it's not that readable, especially as models get larger. Most people who have to actually use ER diagrams seem to use some variation of the UML notation. And even though it doesn't seem to have a name, OPs version is a pretty straightforward (and imho more intuitive) extension of the UML notation (forbidding N-N relationships and adding the relationship name as a foreign key to the table instead of writing it on the arrow). It's also hardly the first time I've seen it, we just need someone to publish a book that gives it a catchy name.

If we want people to actually use ER diagrams, treating the original version from 50 years ago as the only valid way to write them doesn't seem useful. Both science and engineering are about iterative improvement after all.

Looks real nice!

Clean for sure.

First feedback I have is, when you click "Settings" > "Flush Storage" there is no "Hey this button will delete everything you have been working on are you sure you want to do this?" type message to confirm if someone wants to actually flush the storage.

I was about 6 or 7 tables deep messing around with it and was looking at settings and kind of absent mindedly clicked that without thinking and everything was gone :D

Lol no worries. It was pure "I don't know what I expected" for the succeeding 2 seconds.

Other feedback is that it would be nice to be able to set it to say "I want to design a database in PostgreSQL" and have the available datatypes match that. So I can specify postgres specific things like time with timezone, money, point (for geo data) etc.

How do you prefer feedback? Issues in the GH repo?

I tried to have the diagrams be independent of the DBMS that's why types that are not supported in one DBMS get mapped to their closest match in another. For example SET gets mapped to an array of ENUM values in Postgres. So money, point, etc. can be easily added and mapped.

As for the feedback, thank you so much, I really appreciate it.

You can leave it anywhere you find convenient but I think GH would be ideal. It would easier to reference and maybe the community can get involved too :)

This was in the plans when i was just starting out until i realized how much longer it was gonna take just to finish the web app. Can't say anytime soon but i'd love to make this happen

Honestly, what I'm looking for is a tool that generates the docs and diagrams for a database from the database schema itself and lays it out cleanly. So far I've found tools that will try to generate a diagram naively even if there are a hundred tables, and I've found tools that make you recreate your structure twice: once in DDL and the second time in docs/diagramming.

I want Javadoc with scoped diagrams for relational databases.

It's been a while since I looked at SchemaSpy. It has improved a lot since I last looked. Thanks for the recommendation! I wish the layout were a little more polished, but the content/functionality is definitely there.

As long as the relationships are also defined in the DDL, the DDL is the only thing that is needed. You can use an automated diagrammer such as the one in https://visualdb.com/ It shows tables and relationships graphically, and you can click on a table to see its columns and data types.

Javadoc is for Java code. I want something that generates documentation for relational databases, not Java. Hence: Javadoc for relational databases.

You might want to mention that it is a "data modelling tool" somewhere in the description text, as some of your target users will definitely use that term to search for such a solution...

This is awesome I was just looking for something like that. The ability to build the diagram from the SQL source is really cool! Thanks!

At first it was mostly for myself just as a learning experience and a project that would go on my resume. I wasn't even planning to open source it until people started expressing interest in it. Now that it's out and there's feedback we can cater to more specific use cases

Can see it happening for sure. Perhaps porting it to electron and taking in dbms type and host url for the connection will do it

For tools like these, I need the ability to save locally to my own device and not have anything saved/sent to the cloud. draw.io is the best example of this.

Nothing gets sent to the cloud, everything is in your browser in IndexedDB. If you go to Inspect>Application>IndexedDB in the storage section you'll see exactly what gets saved and where

（评论） (comments)

（评论）
(comments)