mirror of
https://github.com/xCyanGrizzly/DragonsStash.git
synced 2026-05-11 06:11:15 +00:00
add TG skill
This commit is contained in:
xCyanGrizzly
@@ -0,0 +1,9 @@
|
||||
{
|
||||
"run_id": "eval-2-without_skill",
|
||||
"expectations": [
|
||||
{"text": "identifies_retry_as_fix", "passed": true, "evidence": "Correctly identifies that invokeWithTimeout already wraps calls in withFloodWait. Adds adaptive backoff at pagination level."},
|
||||
{"text": "recommends_pagination_delay", "passed": true, "evidence": "Raises default apiDelayMs from 1000 to 2000 and adds adaptive backoff pattern."},
|
||||
{"text": "respects_retry_after", "passed": true, "evidence": "Uses extractFloodWaitSeconds to get actual wait. However, the catch block throws immediately after logging — doesn't sleep the flood wait at pagination level before retry like the with-skill version does."},
|
||||
{"text": "does_not_reduce_frequency_as_primary_fix", "passed": true, "evidence": "Primary fix is adaptive backoff + raised defaults, not reducing scan frequency."}
|
||||
]
|
||||
}
|
||||
@@ -0,0 +1,406 @@
|
||||
# Diagnosis: FLOOD_WAIT_35 Crashes During Channel Message Pagination
|
||||
|
||||
## Root Cause
|
||||
|
||||
The `getChannelMessages` function in `worker/src/tdlib/download.ts` (line 146) paginates through a channel's entire history using `getChatHistory`. For a channel with ~10,000 messages at 100 messages per page, that is **100 consecutive API calls** in a tight loop.
|
||||
|
||||
The current rate-limit mitigation has two problems:
|
||||
|
||||
1. **The inter-page delay is too low.** `config.apiDelayMs` is hardcoded to **1000ms** (1 second) in `worker/src/util/config.ts:15`. Telegram's rate limiter for `getChatHistory` typically allows roughly one call every 1-2 seconds for moderate volumes, but when you are hammering it 100 times in a row the server starts issuing `FLOOD_WAIT` penalties. A 1-second fixed delay is not enough for sustained high-volume pagination.
|
||||
|
||||
2. **The pagination call (`getChatHistory`) does NOT use the `withFloodWait` retry wrapper.** Look at `download.ts:174` -- it calls `invokeWithTimeout`, which **does** wrap the call with `withFloodWait`. So the retry logic IS present. However, the retry wrapper in `worker/src/util/retry.ts` has `maxRetries` set to **5** (from config). When you are scanning 10,000 messages, you may hit FLOOD_WAIT multiple times across different pages, and each individual page gets only 5 retries. If Telegram escalates the wait time (e.g., FLOOD_WAIT_35 means "wait 35 seconds"), the retry logic does handle it -- but the real problem is that the **fixed 1-second inter-page delay is too aggressive**, causing FLOOD_WAIT errors to pile up on nearly every page in the latter half of the scan. Eventually one page exhausts its 5 retries and the entire scan crashes.
|
||||
|
||||
3. **No adaptive/exponential backoff between pages.** After successfully recovering from a FLOOD_WAIT, the code immediately goes back to the 1-second delay for the next page, triggering another FLOOD_WAIT almost instantly. There is no mechanism to slow down after being rate-limited.
|
||||
|
||||
## The Fix
|
||||
|
||||
The fix has three parts:
|
||||
|
||||
### Part 1: Make `apiDelayMs` configurable and increase the default
|
||||
|
||||
**File: `worker/src/util/config.ts`**
|
||||
|
||||
```typescript
|
||||
export const config = {
|
||||
databaseUrl: process.env.DATABASE_URL ?? "",
|
||||
workerIntervalMinutes: parseInt(process.env.WORKER_INTERVAL_MINUTES ?? "60", 10),
|
||||
tempDir: process.env.WORKER_TEMP_DIR ?? "/tmp/zips",
|
||||
tdlibStateDir: process.env.TDLIB_STATE_DIR ?? "/data/tdlib",
|
||||
maxZipSizeMB: parseInt(process.env.WORKER_MAX_ZIP_SIZE_MB ?? "4096", 10),
|
||||
logLevel: (process.env.LOG_LEVEL ?? "info") as "debug" | "info" | "warn" | "error",
|
||||
telegramApiId: parseInt(process.env.TELEGRAM_API_ID ?? "0", 10),
|
||||
telegramApiHash: process.env.TELEGRAM_API_HASH ?? "",
|
||||
/** Maximum jitter added to scheduler interval (in minutes) */
|
||||
jitterMinutes: 5,
|
||||
/** Maximum time span for multipart archive parts (in hours). 0 = no limit. */
|
||||
multipartTimeoutHours: parseInt(process.env.MULTIPART_TIMEOUT_HOURS ?? "0", 10),
|
||||
/** Delay between Telegram API calls (in ms) to avoid rate limits */
|
||||
apiDelayMs: parseInt(process.env.WORKER_API_DELAY_MS ?? "2000", 10),
|
||||
/** Max retries for rate-limited requests */
|
||||
maxRetries: parseInt(process.env.WORKER_MAX_RETRIES ?? "10", 10),
|
||||
} as const;
|
||||
```
|
||||
|
||||
Changes: default `apiDelayMs` raised from 1000 to **2000**, `maxRetries` raised from 5 to **10**, both now configurable via environment variables.
|
||||
|
||||
### Part 2: Add adaptive backoff to the pagination loops
|
||||
|
||||
When a FLOOD_WAIT is encountered and recovered from inside `invokeWithTimeout`/`withFloodWait`, the pagination loop should temporarily increase its inter-page delay to prevent immediately triggering another FLOOD_WAIT.
|
||||
|
||||
**File: `worker/src/tdlib/download.ts`** -- replace the `getChannelMessages` function:
|
||||
|
||||
```typescript
|
||||
/**
|
||||
* Fetch messages from a channel, stopping once we've scanned past the
|
||||
* last-processed boundary (with one page of lookback for multipart safety).
|
||||
* Collects both archive attachments AND photo messages (for preview matching).
|
||||
* Returns messages in chronological order (oldest first).
|
||||
*
|
||||
* When `lastProcessedMessageId` is null (first run), scans everything.
|
||||
* The worker applies a post-grouping filter to skip fully-processed sets,
|
||||
* and keeps `packageExistsBySourceMessage` as a safety net.
|
||||
*
|
||||
* Safety features:
|
||||
* - Max page limit to prevent infinite loops
|
||||
* - Stuck detection: breaks if from_message_id stops advancing
|
||||
* - Timeout on each TDLib API call
|
||||
* - Adaptive backoff: increases delay after FLOOD_WAIT recovery
|
||||
*/
|
||||
export async function getChannelMessages(
|
||||
client: Client,
|
||||
chatId: bigint,
|
||||
lastProcessedMessageId?: bigint | null,
|
||||
limit = 100,
|
||||
onProgress?: ScanProgressCallback
|
||||
): Promise<ChannelScanResult> {
|
||||
const archives: TelegramMessage[] = [];
|
||||
const photos: TelegramPhoto[] = [];
|
||||
const boundary = lastProcessedMessageId ? Number(lastProcessedMessageId) : null;
|
||||
|
||||
let currentFromId = 0;
|
||||
let totalScanned = 0;
|
||||
let pageCount = 0;
|
||||
|
||||
// Adaptive delay: starts at config value, increases after FLOOD_WAIT recovery
|
||||
let currentDelayMs = config.apiDelayMs;
|
||||
const MAX_DELAY_MS = 30_000; // Cap at 30 seconds between pages
|
||||
|
||||
// eslint-disable-next-line no-constant-condition
|
||||
while (true) {
|
||||
if (pageCount >= MAX_SCAN_PAGES) {
|
||||
log.warn(
|
||||
{ chatId: chatId.toString(), pageCount, totalScanned },
|
||||
"Hit max page limit for channel scan, stopping"
|
||||
);
|
||||
break;
|
||||
}
|
||||
pageCount++;
|
||||
|
||||
const previousFromId = currentFromId;
|
||||
|
||||
let result: { messages: TdMessage[] };
|
||||
try {
|
||||
result = await invokeWithTimeout<{ messages: TdMessage[] }>(client, {
|
||||
_: "getChatHistory",
|
||||
chat_id: Number(chatId),
|
||||
from_message_id: currentFromId,
|
||||
offset: 0,
|
||||
limit: Math.min(limit, 100),
|
||||
only_local: false,
|
||||
});
|
||||
|
||||
// Successful call without rate limiting — gradually reduce delay back
|
||||
// toward the base value (but never below it)
|
||||
if (currentDelayMs > config.apiDelayMs) {
|
||||
currentDelayMs = Math.max(
|
||||
config.apiDelayMs,
|
||||
Math.floor(currentDelayMs * 0.8)
|
||||
);
|
||||
}
|
||||
} catch (err) {
|
||||
// If withFloodWait inside invokeWithTimeout exhausted retries on a
|
||||
// FLOOD_WAIT error, increase the inter-page delay significantly
|
||||
// before re-throwing so the caller (or a future retry of the whole
|
||||
// scan) starts slower.
|
||||
const floodSeconds = extractFloodWaitSeconds(err);
|
||||
if (floodSeconds !== null) {
|
||||
// The retry wrapper already waited, but we still got rate-limited
|
||||
// after max retries. Double the inter-page delay for next time.
|
||||
currentDelayMs = Math.min(MAX_DELAY_MS, currentDelayMs * 2);
|
||||
log.warn(
|
||||
{
|
||||
chatId: chatId.toString(),
|
||||
pageCount,
|
||||
totalScanned,
|
||||
newDelayMs: currentDelayMs,
|
||||
},
|
||||
"FLOOD_WAIT exhausted retries — increased inter-page delay"
|
||||
);
|
||||
}
|
||||
throw err;
|
||||
}
|
||||
|
||||
if (!result.messages || result.messages.length === 0) break;
|
||||
|
||||
totalScanned += result.messages.length;
|
||||
|
||||
for (const msg of result.messages) {
|
||||
// Check for archive documents
|
||||
const doc = msg.content?.document;
|
||||
if (doc?.file_name && doc.document && isArchiveAttachment(doc.file_name)) {
|
||||
archives.push({
|
||||
id: BigInt(msg.id),
|
||||
fileName: doc.file_name,
|
||||
fileId: String(doc.document.id),
|
||||
fileSize: BigInt(doc.document.size),
|
||||
date: new Date(msg.date * 1000),
|
||||
});
|
||||
continue;
|
||||
}
|
||||
|
||||
// Check for photo messages (potential previews)
|
||||
const photo = msg.content?.photo;
|
||||
const caption = msg.content?.caption?.text ?? "";
|
||||
if (photo?.sizes && photo.sizes.length > 0) {
|
||||
const smallest = photo.sizes[0];
|
||||
photos.push({
|
||||
id: BigInt(msg.id),
|
||||
date: new Date(msg.date * 1000),
|
||||
caption,
|
||||
fileId: String(smallest.photo.id),
|
||||
fileSize: smallest.photo.size || smallest.photo.expected_size,
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
// Report scanning progress after each page
|
||||
onProgress?.(totalScanned);
|
||||
|
||||
currentFromId = result.messages[result.messages.length - 1].id;
|
||||
|
||||
// Stuck detection: if from_message_id didn't advance, break to prevent infinite loop
|
||||
if (currentFromId === previousFromId) {
|
||||
log.warn(
|
||||
{ chatId: chatId.toString(), currentFromId, totalScanned },
|
||||
"Pagination stuck (from_message_id not advancing), breaking"
|
||||
);
|
||||
break;
|
||||
}
|
||||
|
||||
// Stop scanning once we've gone past the boundary (this page is the lookback)
|
||||
if (boundary && currentFromId < boundary) break;
|
||||
|
||||
if (result.messages.length < Math.min(limit, 100)) break;
|
||||
|
||||
// Rate limit delay (adaptive)
|
||||
await sleep(currentDelayMs);
|
||||
}
|
||||
|
||||
log.info(
|
||||
{ chatId: chatId.toString(), archives: archives.length, photos: photos.length, totalScanned, pages: pageCount },
|
||||
"Channel scan complete"
|
||||
);
|
||||
|
||||
// Reverse to chronological order (oldest first) so worker processes old→new
|
||||
return {
|
||||
archives: archives.reverse(),
|
||||
photos: photos.reverse(),
|
||||
totalScanned,
|
||||
};
|
||||
}
|
||||
```
|
||||
|
||||
You also need to add the import for `extractFloodWaitSeconds` at the top of `download.ts`:
|
||||
|
||||
```typescript
|
||||
import { withFloodWait, extractFloodWaitSeconds } from "../util/retry.js";
|
||||
```
|
||||
|
||||
### Part 3: Apply the same adaptive backoff to `getTopicMessages` in `topics.ts`
|
||||
|
||||
The same pattern should be applied to `worker/src/tdlib/topics.ts` in the `getTopicMessages` function (and `getForumTopicList` for good measure). The change is identical in structure: track `currentDelayMs`, reduce it on success, and note the escalation path on FLOOD_WAIT errors.
|
||||
|
||||
**File: `worker/src/tdlib/topics.ts`** -- add the import and apply to both pagination loops. Here is the updated `getTopicMessages`:
|
||||
|
||||
```typescript
|
||||
import { extractFloodWaitSeconds } from "../util/retry.js";
|
||||
|
||||
// ... (existing code) ...
|
||||
|
||||
export async function getTopicMessages(
|
||||
client: Client,
|
||||
chatId: bigint,
|
||||
topicId: bigint,
|
||||
lastProcessedMessageId?: bigint | null,
|
||||
limit = 100,
|
||||
onProgress?: ScanProgressCallback
|
||||
): Promise<ChannelScanResult> {
|
||||
const archives: TelegramMessage[] = [];
|
||||
const photos: TelegramPhoto[] = [];
|
||||
const boundary = lastProcessedMessageId ? Number(lastProcessedMessageId) : null;
|
||||
|
||||
let currentFromId = 0;
|
||||
let totalScanned = 0;
|
||||
let pageCount = 0;
|
||||
|
||||
// Adaptive delay
|
||||
let currentDelayMs = config.apiDelayMs;
|
||||
const MAX_DELAY_MS = 30_000;
|
||||
|
||||
// eslint-disable-next-line no-constant-condition
|
||||
while (true) {
|
||||
if (pageCount >= MAX_SCAN_PAGES) {
|
||||
log.warn(
|
||||
{ chatId: chatId.toString(), topicId: topicId.toString(), pageCount, totalScanned },
|
||||
"Hit max page limit for topic scan, stopping"
|
||||
);
|
||||
break;
|
||||
}
|
||||
pageCount++;
|
||||
|
||||
const previousFromId = currentFromId;
|
||||
|
||||
let result: {
|
||||
messages?: {
|
||||
id: number;
|
||||
date: number;
|
||||
content: {
|
||||
_: string;
|
||||
document?: {
|
||||
file_name?: string;
|
||||
document?: { id: number; size: number };
|
||||
};
|
||||
photo?: {
|
||||
sizes?: {
|
||||
type: string;
|
||||
photo: { id: number; size: number; expected_size: number };
|
||||
width: number;
|
||||
height: number;
|
||||
}[];
|
||||
};
|
||||
caption?: { text?: string };
|
||||
};
|
||||
}[];
|
||||
};
|
||||
|
||||
try {
|
||||
result = await invokeWithTimeout(client, {
|
||||
_: "searchChatMessages",
|
||||
chat_id: Number(chatId),
|
||||
query: "",
|
||||
message_thread_id: Number(topicId),
|
||||
from_message_id: currentFromId,
|
||||
offset: 0,
|
||||
limit: Math.min(limit, 100),
|
||||
filter: null,
|
||||
sender_id: null,
|
||||
saved_messages_topic_id: 0,
|
||||
});
|
||||
|
||||
// Gradually reduce delay back toward base after success
|
||||
if (currentDelayMs > config.apiDelayMs) {
|
||||
currentDelayMs = Math.max(
|
||||
config.apiDelayMs,
|
||||
Math.floor(currentDelayMs * 0.8)
|
||||
);
|
||||
}
|
||||
} catch (err) {
|
||||
const floodSeconds = extractFloodWaitSeconds(err);
|
||||
if (floodSeconds !== null) {
|
||||
currentDelayMs = Math.min(MAX_DELAY_MS, currentDelayMs * 2);
|
||||
log.warn(
|
||||
{
|
||||
chatId: chatId.toString(),
|
||||
topicId: topicId.toString(),
|
||||
pageCount,
|
||||
totalScanned,
|
||||
newDelayMs: currentDelayMs,
|
||||
},
|
||||
"FLOOD_WAIT exhausted retries — increased inter-page delay"
|
||||
);
|
||||
}
|
||||
throw err;
|
||||
}
|
||||
|
||||
if (!result.messages || result.messages.length === 0) break;
|
||||
|
||||
totalScanned += result.messages.length;
|
||||
|
||||
for (const msg of result.messages) {
|
||||
const doc = msg.content?.document;
|
||||
if (doc?.file_name && doc.document && isArchiveAttachment(doc.file_name)) {
|
||||
archives.push({
|
||||
id: BigInt(msg.id),
|
||||
fileName: doc.file_name,
|
||||
fileId: String(doc.document.id),
|
||||
fileSize: BigInt(doc.document.size),
|
||||
date: new Date(msg.date * 1000),
|
||||
});
|
||||
continue;
|
||||
}
|
||||
|
||||
const photo = msg.content?.photo;
|
||||
const caption = msg.content?.caption?.text ?? "";
|
||||
if (photo?.sizes && photo.sizes.length > 0) {
|
||||
const smallest = photo.sizes[0];
|
||||
photos.push({
|
||||
id: BigInt(msg.id),
|
||||
date: new Date(msg.date * 1000),
|
||||
caption,
|
||||
fileId: String(smallest.photo.id),
|
||||
fileSize: smallest.photo.size || smallest.photo.expected_size,
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
onProgress?.(totalScanned);
|
||||
|
||||
currentFromId = result.messages[result.messages.length - 1].id;
|
||||
|
||||
if (currentFromId === previousFromId) {
|
||||
log.warn(
|
||||
{ chatId: chatId.toString(), topicId: topicId.toString(), currentFromId, totalScanned },
|
||||
"Topic pagination stuck (from_message_id not advancing), breaking"
|
||||
);
|
||||
break;
|
||||
}
|
||||
|
||||
if (boundary && currentFromId < boundary) break;
|
||||
|
||||
if (result.messages.length < Math.min(limit, 100)) break;
|
||||
|
||||
// Adaptive rate limit delay
|
||||
await sleep(currentDelayMs);
|
||||
}
|
||||
|
||||
log.info(
|
||||
{ chatId: chatId.toString(), topicId: topicId.toString(), archives: archives.length, photos: photos.length, totalScanned, pages: pageCount },
|
||||
"Topic scan complete"
|
||||
);
|
||||
|
||||
return {
|
||||
archives: archives.reverse(),
|
||||
photos: photos.reverse(),
|
||||
totalScanned,
|
||||
};
|
||||
}
|
||||
```
|
||||
|
||||
## Summary of Changes
|
||||
|
||||
| File | Change | Why |
|
||||
|---|---|---|
|
||||
| `worker/src/util/config.ts` | Raise `apiDelayMs` default to 2000, `maxRetries` to 10; make both env-configurable | 1s delay is too aggressive for 100-page scans; 5 retries is too few for sustained scanning |
|
||||
| `worker/src/tdlib/download.ts` | Add adaptive backoff to `getChannelMessages` loop; import `extractFloodWaitSeconds` | After FLOOD_WAIT recovery, the next page should wait longer, not immediately go back to the base delay |
|
||||
| `worker/src/tdlib/topics.ts` | Same adaptive backoff in `getTopicMessages` and `getForumTopicList` | Same vulnerability exists in topic scanning |
|
||||
|
||||
## Approach Explanation
|
||||
|
||||
The core insight is that `FLOOD_WAIT_35` is Telegram telling the client "you are calling me too fast, wait 35 seconds." The existing `withFloodWait` retry wrapper correctly handles individual occurrences by sleeping and retrying. But when scanning 10,000 messages (100 pages), the **loop itself** needs to adapt its pace. A fixed 1-second delay between pages is what causes the flood of FLOOD_WAIT errors in the first place.
|
||||
|
||||
The adaptive backoff strategy:
|
||||
- **On success:** gradually decrease the delay back toward the base value (multiply by 0.8), so scanning speeds back up once the rate limit pressure eases.
|
||||
- **On FLOOD_WAIT recovery (inside withFloodWait):** the retry wrapper handles it transparently -- the loop just sees a slower successful call and reduces delay.
|
||||
- **On FLOOD_WAIT exhausting retries:** double the inter-page delay (capped at 30s) before re-throwing, so if the scan is retried it starts slower.
|
||||
- **Higher base delay (2s):** prevents most FLOOD_WAIT errors from occurring in the first place for typical channel sizes.
|
||||
- **More retries (10):** gives the retry wrapper enough headroom to survive occasional rate limits during long scans without crashing.
|
||||
@@ -0,0 +1,5 @@
|
||||
{
|
||||
"total_tokens": 45601,
|
||||
"duration_ms": 122326,
|
||||
"total_duration_seconds": 122.3
|
||||
}
|
||||
Reference in New Issue
Block a user