Cross-domain picture anti-blocking asynchronous processing method and system based on time sequence supervised learning

Abstract

The application discloses a cross-domain picture anti-blocking asynchronous processing method and system based on time sequence supervised learning, relates to the technical field of electric digital data processing, and has the advantages that through hierarchical weighted operation, a dynamic generation of a fragmentation optimistic waiting time length is realized, regional time delay adaptive correction is carried out, fixed time length misjudgment is avoided to generate a residual code stream, the business base of a consumer partition thread spin waiting is reduced, thread resource invalid occupation is relieved, a time sequence network parameter is used to train a blocking prediction model, a thread pool blocking portrait is iteratively updated to link a waiting time length, a waiting window is dynamically adjusted, a high-risk session is terminated in advance, blocking threads are prevented from being long-term resident and being converted into zombie threads, and the proportion of active threads is improved; meanwhile, global and merchant blocking features are quantified to generate dynamic routing offset parameters, tasks are evenly distributed to avoid partition blocking storm, multiple mechanisms are cooperatively used to reduce the total amount of global spin threads, zombie threads are inhibited from occupying resources, and the stability of commodity picture asynchronous processing concurrent scheduling is comprehensively improved.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, specifically a cross-domain image anti-blocking asynchronous processing method and system based on time-series supervised learning. Background Technology

[0002] In cross-region asynchronous synchronization of product images, conventional solutions set a fixed sharding wait time, which cannot adapt to the network latency differences between different storage regions, easily resulting in a large number of incomplete bitstreams. Message distribution relies on fixed hash routing, and tasks carrying incomplete bitstreams will flood the same consumer partition. Within this partition, a large number of threads continuously spin, waiting for resource release, gradually forming zombie threads that consume system computing power and memory, ultimately causing continuous blockage of the entire asynchronous product image processing chain. Therefore, there is an urgent need for a cross-domain image anti-blocking asynchronous processing method and system based on time-series supervised learning. Summary of the Invention

[0003] The purpose of this invention is to provide a cross-domain image anti-blocking asynchronous processing method and system based on temporal supervised learning, so as to solve the problems raised in the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a cross-domain image anti-blocking asynchronous processing method based on temporal supervised learning, the cross-domain image anti-blocking asynchronous processing method comprising the following steps: Step S1: Deploy a synchronization session lifecycle event listener in the cross-regional synchronization scheduling service layer, select the data observation interface of the distributed KV storage, and perform continuous event listening on the entire lifecycle status of the asynchronous cross-regional synchronization of product images in all data areas. The data area is represented as a business storage area with an independently deployed distributed storage cluster, a dedicated cross-regional transmission line, and a globally unique regional identifier. Simultaneously establish a sharded optimistic waiting window timing mechanism, and independently bind the corresponding waiting timing period to each cross-regional synchronization session. Distributed KV storage refers to a distributed key-value cache database. Mainstream engineering products include RedisCluster and Tair. The core storage logic is that each piece of data consists of a unique, non-repeating key retrieval primary key and a corresponding value business load. Data reading and writing, expiration monitoring, and metadata collection are completed only through the primary key. The read and write latency is extremely low, and it supports cross-data center multi-replica synchronization. It is specifically designed to carry lightweight data with high-frequency read and write frequency and short lifespan, such as session temporary states, intermediate variables in calculations, and sharding time markers. In the storage primary key concatenation rules, the format is {region identifier, session ID, business type imagetransfer}. Iterate through and parse the historical operation logs of each data area, extract all historical synchronization session duration data of the local data area and all historical synchronization session duration data of each remote data area; calculate the arithmetic mean of all local session durations to generate the first local baseline session waiting time; calculate the arithmetic mean of all remote data area session durations to generate the comprehensive session waiting time of the entire data area. The difference between the local session duration and the corresponding remote regional session duration within each historical session is calculated, and all differences are summarized to form a difference sample set. The arithmetic mean of the difference sample set is calculated to obtain the regional session duration deviation feature. The regional comprehensive session waiting time of the entire data area is selected as the basis and superimposed with the regional session duration deviation feature to calculate the second local benchmark session waiting time. The total number of local session samples and the total number of remote session samples in the historical logs are statistically analyzed, and the ratio of the two types of samples is selected as the duration fusion weight. This duration fusion weight is used to perform a weighted fusion operation on the first local baseline session waiting time and the second local baseline session waiting time. The weighted fusion operation output is the current optimistic waiting time of the fragment that is adapted to the local data area network baseline. By default, the whole-domain fragment transmission process is a complete transmission state, and only the regional baseline deviation is selected to complete the duration adaptive correction, so as to achieve accurate local adaptation of the fragment waiting time. Existing technologies generally use fixed fragmentation waiting time, which does not distinguish the network baseline differences of different data areas, and the transmission latency deviation between local data areas and global data areas cannot be quantitatively adapted; if the fixed waiting time is too short, normal slow transmission will be misjudged as fragmentation missing, and the session will be terminated prematurely, resulting in incomplete bit streams; if the fixed waiting time is too long, it will cause long-term occupation of effective session resources, session accumulation, and increased frequency of invalid polling. By monitoring the global session status and parsing historical logs to calculate two types of local baseline durations, the session sample size is selected to generate fusion weights to complete the weighted calculation. The optimistic waiting duration of the shards is dynamically calculated to adapt to the local network baseline, and the waiting judgment criteria for the sharding and aggregation of data regions are unified to eliminate the misjudgment and session resource waste caused by the fixed duration mechanism. Step S2: Deploy dedicated network feature monitoring points in the session context instance container corresponding to each synchronization session, collect the link operation status of historical synchronization sessions in the local data area in full time series, and continuously extract four types of time series network parameters: instantaneous synchronization throughput, average transmission rate per minute, dedicated line packet loss rate, and TCP retransmission count. Four types of temporal network parameters were selected as model input features. The fragmented state generated by the actual historical session, the spin time of KV lock, the number of task queues backlog, and the number of fragmented bitstreams generated were selected as sample annotation data to construct a temporal training sample set. The temporal training sample set was subjected to standardization preprocessing and temporal alignment processing. The model was iteratively converged by temporal supervised training and solidified to obtain the image asynchronous processing blocking prediction model. Within the current session fragmentation optimistic waiting time duration, the instantaneous synchronous throughput, average transmission rate, leased line packet loss rate, and TCP retransmission count of the current cross-regional link are collected in real time through context feature monitoring points. The four types of network parameters collected in real time are synchronously input into the image asynchronous processing blocking prediction model. Through time-series feature inference calculation, four types of structured prediction data are uniformly output: the probability value of incomplete generation, the expected total spin time of KV batch read lock, the expected length of the synchronous task queue, and the expected incremental value of incomplete code stream messages within a future preset time period. Existing technologies rely solely on static network thresholds to determine link status, making it impossible to quantitatively predict fragmentation trends, lock spin consumption, task backlog, and fragmented bitstream output. Instantaneous network fluctuations cannot be captured in a timely manner, and delayed perception of link anomalies leads to a large number of fragmented fragments entering the merging process, inducing chronic thread blocking problems. By selecting dedicated listening points for the session context, we can achieve refined time-series collection of network parameters. By selecting real business samples to complete supervised training of the model, we can realize pre-quantitative prediction of abnormal link states and transform instantaneous network characteristics into calculable and quantifiable congestion risk indicators. Step S3: Parse the historical session running logs of the local data area, extract the thread pool running status data in a structured manner, and distinguish between active worker threads and blocked worker threads in the asynchronous image processing process; among them, blocked worker threads are active worker threads that occupy running resources for a long time due to distributed lock spinning, decoding loop exceptions and link wait exceptions. If blocked worker threads cannot release resources, they will gradually turn into zombie threads. The proportion of active worker threads, the increase in the number of blocked worker threads, the cumulative number of zombie threads, and the duration of thread starvation within the historical session period are statistically analyzed to form a multi-dimensional basic feature dataset characterizing the thread pool's operating state. Zombie threads continuously occupy the thread pool's resident resources, leading to the depletion of idle threads and the inability to schedule new tasks. They also continuously hold file handles, distributed read-write locks, and network connections, causing persistent memory and computing power consumption. This continuously increases lock contention and retransmission overhead, forming a chain-like correlation feature of link fluctuations, incomplete bitstreams, thread deadlocks, and synchronous blocking. Four predictive metrics are selected from the output of step S2: incomplete generation probability value, expected total spin time of batch KV read lock, expected task queue backlog length, and expected incomplete code stream message increment. These four predictive metrics are then matched with the thread pool's multi-dimensional basic feature dataset. Using these four predictive metrics as the basis for matching and retrieval, all historical thread basic feature samples are traversed, and historical samples with similarity are selected. The active worker thread occupancy ratio, the number of blocked worker threads, the cumulative number of zombie threads, and the duration of thread starvation corresponding to the matched samples are extracted. The mean of the parameters of the selected samples is calculated to complete parameter aggregation, resulting in an ordered numerical sequence composed of the four means. This ordered numerical sequence is used as a feature vector, and a thread pool blocking profile is constructed based on the feature vector in the local data region. Existing technologies only count the number of threads, which cannot correlate network link anomalies with thread blocking, nor can they quantify the extent to which incomplete bitstreams erode thread pool resources. The thread pool's running status can only be observed after the fact, and cannot form a structured, comparable, and computable feature profile system.

[0005] This invention achieves deep integration of network prediction features and thread operation features, solidifies the implicit thread blocking evolution trend, resource leakage trend, and link blocking propagation trend into a standardized thread pool blocking profile, and realizes the quantitative representation of the thread pool operation status in the local data area. Step S4: Read all historical thread pool blocking profiles that are persistently stored in the local data area, aggregate them to form a complete historical profile set, and compare the thread pool blocking profiles that were built in real time in Step S3 with each historical profile in the historical profile set in turn to perform independent difference comparisons of four feature dimensions. The four feature dimensions are the proportion of active worker threads, the number of blocked worker threads, the cumulative number of zombie threads, and the duration of thread starvation. Quantitative statistical calculations of feature weights are performed for each of the four feature dimensions. All portrait samples in the historical portrait set are traversed. For a single feature dimension, the dispersion of the feature value of that single dimension in the sample set is calculated. At the same time, the cumulative changes of three types of business abnormal data, namely thread blocking increment, incomplete code stream generation increment, and thread resource leakage increment, are calculated when the feature value of that dimension fluctuates. The cumulative changes in thread blocking, incomplete bitstream, and thread resource leakage corresponding to a single dimension are summed to obtain the cumulative value of anomaly association for that dimension feature. The cumulative values ​​of anomaly association for each of the four feature dimensions are calculated sequentially, and the cumulative values ​​of the four anomaly associations are summed to obtain the total global anomaly. The ratio of the cumulative value of anomaly association for a single dimension to the total global anomaly is used as the final feature weight for that dimension. Based on the above quantitative statistical method, the following weights are generated: active worker thread occupancy ratio weight, blocked worker thread growth weight, zombie thread cumulative number weight, and thread starvation duration weight. The weight values ​​of each dimension are uniquely determined by the proportion of abnormal correlations in historical sample data. By incorporating the feature weights obtained from quantitative calculations across various dimensions, and combining them with the four-dimensional feature differences between the real-time image and each historical image, the global matching degree value of the real-time image relative to each historical image is calculated for each image. After completing the matching calculation for all historical image samples, the historical image sample with the highest global matching degree value is selected. The formula for calculating the global matching degree value is as follows: ; In the formula, The distance between the real-time profile and the historical profile is represented by Mahalanobis distance. The smaller the distance, the higher the similarity of the thread pool's running state. This is represented as a real-time constructed four-dimensional thread pool blocking feature vector. ,include; This represents the average percentage of locally active worker threads. This represents the average increase in the number of local real-time blocking worker threads. This represents the average cumulative number of zombie threads in the local real-time context. This is expressed as the average duration of local real-time thread starvation. Represented as a historically constructed four-dimensional thread pool blocking feature vector, including parameters and same; This is represented as a vector transpose operation. By calculating the covariance matrix of the four-dimensional thread features, the coupling relationship of the features in each dimension can be obtained. It is represented as the inverse covariance matrix and is used to eliminate interference from characteristic dimensions and linear correlation. The historical best shard waiting time parameter bound to the historical profile with the highest matching degree is read, and the incomplete generation probability value output in step S2 is retrieved in real time. Using the historical best shard waiting time as the calculation base, the product of the historical best shard waiting time and the incomplete generation probability value is used as the correction compensation amount. The new shard waiting time value is obtained by superimposing the correction compensation amount on the base, and the new shard waiting time value is directly overwritten with the optimistic waiting time of the current shard output in step S1. The calculation formula for the optimistic waiting time of the current shard is as follows: ; In the formula, This is represented as the new fragment waiting time value obtained by adjusting the compensation amount through base addition; This is expressed as the optimal duration that minimizes the loss function within the constrained interval of the upper and lower limits of the fragment duration. The table shows the minimum and maximum waiting times for each partition, as obtained from system statistics. It is represented as a joint loss function, used to uniformly quantify the service loss of incomplete bitstream and the computing resource loss of threads; This is represented as the weighting coefficient for the incomplete fragment anomaly loss; This represents the probability of predicting the incomplete bitstream in the link when the waiting time is T. This is represented as the weighting coefficient for thread spin resource consumption loss; This is expressed as the mathematical expectation of the spin time of a KV distributed lock; Step S5: Based on the optimal optimistic waiting time of the overwritten shards, continuously monitor the thread pool operation characteristics and link operation characteristics within each waiting window, and statistically analyze the frequency of residual shard generation, the cumulative number of spin-blocking threads, and the duration of continuous task queue accumulation within each window. Integrate the three statistical results to obtain the residual bitstream blocking characteristic parameters of the local data area; synchronously collect thread pool operation data and residual bitstream generation data of all remote data areas, and statistically analyze the frequency of residual shard generation, the cumulative number of spin-blocking threads, and the duration of continuous task queue accumulation in the corresponding remote areas, and independently calculate the residual bitstream blocking characteristic parameters of each remote data area. We perform weighted aggregation calculations on the incomplete bitstream blocking feature parameters of the local data area and all remote data areas to obtain the comprehensive incomplete bitstream blocking feature parameters of the entire domain; we perform data isolation and aggregation on the cross-regional asynchronous image processing tasks of the entire domain according to the merchant dimension, and respectively count the cross-regional synchronous task volume, incomplete bitstream generation quantity, thread blocking frequency, and task backlog duration of each merchant, and integrate the calculations to obtain the comprehensive incomplete bitstream blocking feature parameters of each merchant. The integrated calculation involves performing global extreme value normalization on the total number of cross-regional synchronization tasks for a single merchant, the total number of incomplete code streams generated, the frequency of thread blocking triggers, and the average backlog time of the task queue, and mapping the four types of indicators to the same numerical range. Based on the historical abnormal contribution ratio of the four types of indicators to the merchant's blocking status, the calculation ratio of the normalized indicators is allocated by using fixed statistical weights; the four normalized and weighted indicators are summed to uniquely output the merchant's comprehensive incomplete bitstream blocking characteristic parameters. Based on the numerical differences in the comprehensive incomplete bitstream blocking characteristic parameters of each merchant, differentiated message distribution offset parameters are generated; hash route offset calculation is performed on the asynchronous image processing messages generated after the fragmentation and merging are completed, and the message route offset is dynamically adjusted according to the merchant blocking characteristics. Merchant messages corresponding to high blocking characteristics are evenly distributed to different consumption partitions in the whole domain, and merchant messages corresponding to low blocking characteristics are neatly filled into idle consumption partitions. Existing technologies use a fixed hash routing method to distribute asynchronous messages. Merchants' high-blocking tasks and incomplete message streams are easily concentrated in a single consumer partition, causing single-partition threads to spin in batches, zombie threads to proliferate, partition task skew, and increased blocking of global asynchronous processing. Multiple merchant tasks interfere with each other and resource contention is severe.

[0006] This step achieves balanced global distribution of asynchronous tasks for merchants with different blocking levels through global multi-region blocking quantification, merchant-level blocking aggregation, and dynamic route offset adjustment. It completely eliminates the partition task skew problem caused by fixed hash distribution, blocks the partition thread blocking storm induced by the concentration of incomplete bitstreams, and achieves stable parallel scheduling of asynchronous image processing tasks for multiple merchants across regions. Furthermore, a cross-domain image anti-blocking asynchronous processing system based on temporal supervised learning is provided, which includes a temporal control module, a risk prediction module, a profile construction module, and a routing scheduling module. The timing control module is used to collect synchronous session data, calculate and iteratively update the optimistic waiting time of the shards; the risk prediction module is used to collect link timing network parameters, train the blocking prediction model and output structured prediction data; the profile construction module is used to extract thread pool running features, construct thread pool blocking profiles and calculate feature weights; the routing scheduling module is used to statistically analyze the blocking feature parameters of the entire domain and merchant incomplete code streams and control the routing partitions of asynchronous image processing messages. The timing control module includes a session monitoring unit and a duration calculation unit. The session monitoring unit is used to deploy a synchronous session lifecycle monitoring component, collect cross-regional synchronous session full-cycle data of product images in each data region, and establish a sharded optimistic waiting window timing mechanism. The duration calculation unit is used to parse the global historical session logs, calculate two types of local baseline waiting durations in a hierarchical manner, weightedly fuse and output the sharded optimistic waiting duration, and link the thread pool blocking profile to iteratively update the sharded optimistic waiting duration. The risk prediction module includes a link acquisition unit and a model inference unit. The link acquisition unit is used to deploy network feature monitoring points in the session context instance container and continuously collect link time-series network parameters. The model inference unit is used to construct a time-series training sample set to complete model training and output multiple types of structured blocking prediction data based on real-time link parameters. The profile construction module includes a thread parsing unit and a feature matching unit. The thread parsing unit is used to read historical session running logs, extract thread pool running metrics, and generate a multi-dimensional basic feature dataset of the thread pool. The feature matching unit is used to retrieve and match historical thread samples to aggregate and generate feature vectors, construct a thread pool blocking profile in the local data area, quantitatively calculate the weight of each feature dimension, and complete the profile set matching retrieval. The routing scheduling module includes a global statistics unit and a message distribution unit. The global statistics unit is used to collect incomplete bitstream blocking feature parameters for each data region and aggregate merchant-level indicators to generate comprehensive incomplete bitstream blocking feature parameters for merchants. The message distribution unit is used to select comprehensive incomplete bitstream blocking feature parameters for merchants to generate message distribution offset parameters and perform hash routing offset calculations to adjust the asynchronous image processing message delivery and consumption partitions.

[0007] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention dynamically generates optimistic waiting times for fragments through layered weighted calculations of the entire log, and adaptively corrects the latency of the region where the fragment is fully transmitted by default, avoiding the problem of generating incomplete bitstreams due to misjudgment of fixed duration. This reduces the inflow of incomplete fragments into the message queue from the source, significantly reducing the number of business operations that trigger thread spin-waiting in the consumption partition, reducing the probability of generating long-term lock-holding and blocking threads, and alleviating the situation of ineffective thread resource occupation.

[0008] 2. This invention trains a blocking prediction model based on temporal network parameters, pre-quantifies and predicts the probability of incomplete generation, lock spin time, and task backlog, and iteratively updates the sharding waiting time in conjunction with historical thread pool blocking profiles. It accurately matches link blocking risks, dynamically adjusts the waiting window, terminates high-risk synchronization sessions in advance, prevents blocked threads from becoming zombie threads due to prolonged resident status, and increases the proportion of effectively active threads in the thread pool.

[0009] 3. This invention quantifies the blocking characteristics of incomplete bitstreams at both the global and merchant levels, generating dynamic message distribution offset parameters to evenly distribute tasks to each consumption partition, thus avoiding the accumulation of incomplete tasks that could cause a partition thread blocking storm. Combined with adaptive sharding wait time management, the total number of global spinning threads continuously decreases, the problem of zombie thread resource encroachment is suppressed, and the overall stability of asynchronous processing and concurrent scheduling of product images is significantly improved. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating a cross-domain image anti-blocking asynchronous processing method based on temporal supervised learning according to the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] Example 1: As Figure 1As shown, this invention provides a technical solution: a cross-domain image anti-blocking asynchronous processing method based on temporal supervised learning. The cross-domain image anti-blocking asynchronous processing method includes the following steps: Deploy a session lifecycle monitoring mechanism to collect historical session logs from the entire domain and calculate the sharded optimistic wait time through hierarchical weighted fusion. Based on the session context feature monitoring, time-series network parameters are collected to complete feature extraction and model training, and link blocking risk status is predicted in real time. Set up a synchronization session lifecycle monitoring component, call the distributed KV storage data observation interface to collect the full lifecycle data of cross-region synchronization sessions of images in each data area. The data area is a business storage area equipped with an independent storage cluster, a dedicated transmission line and a unique area identifier. Allocate an independent sharding waiting timer period for each synchronization session. The historical logs of each region are analyzed to separate the local data region and the synchronization session duration data of the other data regions. The session mean of the local data region and the session mean of all other data regions are calculated respectively. The average session duration difference is calculated in turn to obtain the regional latency deviation characteristics. The second local baseline waiting time is generated by combining the session mean of all other data regions. The ratio of the total number of session samples in the local data region to all other data regions is used as the fusion weight. The two types of local baseline waiting times are weighted and the output is the segmented optimistic waiting time. The waiting time is then corrected by combining the regional latency deviation to obtain the segmented optimistic waiting time. In specific implementation, taking the cross-regional asynchronous synchronization business of product images on an e-commerce platform as an example, during the execution of this step, the system deploys the SyncSessionLifecycleEventListener component in the cross-regional synchronization scheduling service layer, connects to the distributed KV storage data observation interface to complete the continuous monitoring of the synchronization sessions in each independent storage area, and the synchronization session lifecycle monitoring component allocates an independent segment waiting timer period for each cross-regional synchronization session corresponding to the transmission of each material segment. The system reads the global historical logs to split the local and remote area session duration data, calculates the two types of baseline waiting durations in layers and extracts the regional latency deviation characteristics, generates fusion weights based on the proportion of global session samples to complete the adaptive correction of segment duration, cancels the global fixed duration configuration logic, and adapts the segment aggregation judgment rules to the dedicated line transmission conditions in each area, eliminating the adverse phenomena of incomplete segment error generation and idle sessions occupying computing resources for a long time in low-speed network scenarios; Based on the session context feature monitoring, time-series network parameters are collected to complete feature extraction and model training, and link blocking risk status is predicted in real time. Based on the session context instance container corresponding to each synchronization session, set network feature monitoring points, collect complete time-series operation data of the historical synchronization sessions in the local data area, and extract time-series network parameters such as instantaneous synchronization throughput, average transmission rate, leased line packet loss rate, and TCP retransmission count. Four types of temporal network parameters are selected as model input features. The fragmented state of historical sessions, KV lock spin time, task queue backlog, and fragmented bitstream generation are selected as sample annotation data to construct a temporal training sample set. Standardization preprocessing and temporal alignment processing are performed on the temporal training sample set. The model is iteratively converged through temporal supervised training and the model is solidified and stored. The model is an asynchronous image processing blocking prediction model. Within the time-limited period corresponding to the optimistic waiting time of the segment, the four types of time-series network parameters of the cross-regional link are collected in real time based on the network feature monitoring points and input into the image asynchronous processing blocking prediction model. After time-series feature inference and calculation, the structured prediction data of the incomplete generation probability value, expected total spin time of KV batch read lock, expected synchronous task queue accumulation length and expected incomplete code stream message increment corresponding to the future preset time period are output to obtain the real-time predicted link blocking risk status. In practical implementation, the session context container carrying a single batch of transmission tasks is equipped with independent network feature monitoring points to continuously collect four types of time-series link parameters during the cross-regional transmission of fragments. Combined with fragment incompleteness and lock spin-related annotation data retained from historical synchronous services, a time-series training sample set is constructed. After time-series supervised training, the asynchronous processing blocking prediction model is converged and solidified. The synchronous session lifecycle monitoring component locks the optimistic waiting period corresponding to the current fragment and starts the model to perform real-time simulation. It outputs multiple blocking quantitative indicators for subsequent periods of the link. The transmission risks caused by instantaneous network fluctuations can be quantitatively determined in advance, making up for the inherent shortcomings of static thresholds in predicting fragment incompleteness trends. Analyze historical session thread execution data, integrate model prediction metrics, and construct a thread pool blocking profile for the local data region. Based on the session context instance container corresponding to each synchronization session, set network feature monitoring points, collect complete time-series operation data of the historical synchronization sessions in the local data area, and extract time-series network parameters such as instantaneous synchronization throughput, average transmission rate, leased line packet loss rate, and TCP retransmission count. Four types of temporal network parameters were selected as model input features. The fragmented state of historical sessions, KV lock spin time, task queue backlog, and fragmented bitstream generation were selected as sample annotation data to construct a temporal training sample set. Standardization preprocessing and temporal alignment processing were performed on the temporal training sample set. The model iterative convergence was achieved through temporal supervised training, and the image asynchronous processing blocking prediction model was solidified and stored. Within the time-limited period corresponding to the optimistic waiting time of the segment, the four types of time-series network parameters of the cross-regional link are collected in real time based on the network feature monitoring points and input into the image asynchronous processing blocking prediction model. After time-series feature inference and calculation, the structured prediction data of the incomplete generation probability value, expected total spin time of KV batch read lock, expected synchronous task queue accumulation length and expected incomplete code stream message increment corresponding to the future preset time period are output. In practical implementation, taking the cross-regional asynchronous synchronization business of e-commerce platform products as an example, the system analyzes and stores the historical thread running logs of the full synchronization process, distinguishes between active threads responsible for material decoding and segmentation and blocked threads caused by link waiting and distributed lock contention, sorts out the complete evolution process of blocked threads gradually turning into zombie threads, builds a multi-dimensional thread feature dataset, retrieves and matches historical thread samples by linking the segmentation blocking prediction index output by the synchronization session lifecycle monitoring component and the model, aggregates the mean values ​​of the selected multiple sets of thread features to generate a four-dimensional feature vector, builds a blocking profile of the thread pool in the local storage area based on the feature vector, and builds a complete quantitative correlation system between transmission link anomalies and backend thread resource consumption; Based on the quantitative statistical analysis of historical portrait samples, feature weights are adaptively calculated to complete the matching and retrieval of real-time portraits and historical portrait sets. The optimal segmentation optimistic waiting time is iteratively updated by combining the probability of incomplete generation. Read all historical thread pool blocking profiles persistently stored in the local data area to form a historical profile set. Then, perform a four-dimensional feature difference comparison between the real-time generated thread pool blocking profile and all profiles in the historical profile set. The four-dimensional features are the active worker thread occupancy ratio, the number of blocked worker threads, the cumulative number of zombie threads, and the duration of thread starvation. Traverse all samples in the historical portrait set, and for a single feature dimension, calculate the numerical dispersion and cumulative changes of thread blocking increment, incomplete bit stream generation increment, and thread resource leakage increment generated by synchronization. Accumulate the three types of increments to obtain the cumulative value of single-dimensional anomaly association. Determine the feature weight of each dimension based on the proportion of the cumulative value of all-dimensional anomaly association. Substitute the feature weights of each dimension to calculate the global matching degree between the real-time image and each historical image, filter the historical image with the highest global matching degree, and read the historical best shard waiting time parameter bound to the image. The model's real-time output of the probability value of incomplete generation is retrieved. Using the historical best sharding waiting time as the base, the product of the base and the probability value of incomplete generation is added to obtain a new sharding waiting time value. The new sharding waiting time value is used to cover the original optimistic sharding waiting time. Two sets of mutually exclusive judgment logic are set up: when the probability value of incomplete generation is lower than the fixed statistical threshold and the cumulative number of zombie threads in the corresponding historical profile is zero, the historical best shard waiting time is retained; when the probability value of incomplete generation is greater than or equal to the fixed statistical threshold or the cumulative number of zombie threads in the corresponding historical profile is greater than zero, the optimistic waiting time of the current shard is updated by numerical superposition. In practical implementation, all thread pool blocking profiles in local persistent storage are read to form a complete sample set. The real-time blocking profile generated by the current batch transmission is compared with all historical profiles using four-dimensional feature difference. The cumulative increment of incomplete fragments, thread blocking, and resource leakage associated with each feature dimension is calculated to complete the autonomous solution of feature weights. The synchronous session lifecycle monitoring component retrieves the baseline fragment duration bound to the optimal matching historical profile and the incomplete generation probability output by the model. The optimistic waiting time of the fragment is iteratively updated by combining the Mahalanobis distance matching rule and the multi-objective convex optimization algorithm. With the help of two sets of mutual exclusion judgment logic, the original duration parameter is selected to be retained or overwritten according to the system running status to complete the dynamic self-correction of the timing window of all synchronous sessions. Full-domain multi-regional blocking quantitative statistics, merchant-level blocking feature aggregation calculation of merchant comprehensive incomplete bitstream blocking feature parameters, generating differentiated message distribution offset parameters to regulate asynchronous image processing message routing partitions; The optimistic waiting time of the updated shard is selected as the benchmark to continuously collect the thread pool operation characteristics and link operation characteristics within each waiting window. The frequency of incomplete shard generation, the cumulative number of spin-blocking threads, and the duration of continuous task queue accumulation within the window are statistically analyzed. The incomplete bitstream blocking characteristic parameters of the local data area are obtained through integrated calculation. At the same time, the thread pool operation data and incomplete bitstream generation data of all other data areas are collected. The three types of indicators of the corresponding areas are statistically analyzed and independently calculated to obtain the incomplete bitstream blocking characteristic parameters of the other data areas. A weighted aggregation operation is performed on the incomplete bitstream blocking feature parameters of the local data area and all other data areas to generate comprehensive incomplete bitstream blocking feature parameters of the entire domain; the merchant's unique identifier is selected as the grouping condition to isolate and collect cross-region asynchronous processing tasks of product images across the entire domain, and the total number of cross-region synchronous tasks, the total number of incomplete bitstreams generated, the frequency of thread blocking triggers, and the average backlog time of the task queue are statistically analyzed for a single merchant. Four statistical indicators for a single merchant are selected and subjected to global extreme value normalization in sequence to map all indicators to a unified value range; fixed statistical weights are assigned to the historical abnormal contribution ratios of the four indicators, and the four normalized and weighted indicators are summed to output the merchant's comprehensive incomplete bitstream blocking feature parameters. The numerical range of the comprehensive incomplete bitstream blocking feature parameters of each merchant is selected to generate the corresponding message distribution offset parameters. The hash route offset operation is performed on the asynchronous processing message of the product image generated by the shard merging. The routing offset is dynamically adjusted according to the blocking feature of the merchant. The merchant message corresponding to the high blocking feature is assigned to multiple consumption partitions in the whole domain, and the merchant message corresponding to the low blocking feature is assigned to the idle consumption partition in the whole domain. In practical implementation, the optimal optimistic waiting time for the adapted fragment transmission is used as the unified statistical benchmark for the entire domain. Thread pool and link operation data during the material synchronization process in each storage area are continuously collected. The frequency of incomplete fragment generation, cumulative scale of spinning threads, and task queue backlog duration within the window are integrated to calculate the single-area blocking characteristic parameters. The parameters of each area in the entire domain are weighted and aggregated to obtain the overall blocking characteristic parameters of the entire domain. SyncSessionLifecycleEventListener classifies and collects the synchronization tasks of the entire domain according to the unique identifier of the merchant. After performing global extreme value normalization on the four types of operation indicators of a single merchant, the merchant blocking characteristic parameters are obtained by weighted summation. Based on the parameter range, differentiated hash routing offset parameters are generated to adjust the asynchronous message distribution path carrying fragments. The synchronization tasks of high-blocking merchants are distributed to multiple groups of consumption partitions, and the tasks of low-blocking merchants fill idle partitions, eliminating the problem of concentrated backlog of incomplete tasks in a single partition and batch thread spinning blocking caused by fixed hash distribution.

[0013] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within the present invention.

Claims

1. A cross-domain image anti-blocking asynchronous processing method based on temporal supervised learning, characterized in that: The cross-domain image anti-blocking asynchronous processing method includes: Deploy a session lifecycle monitoring mechanism to collect historical session logs from the entire domain and calculate the sharded optimistic wait time through hierarchical weighted fusion. Based on the session context feature monitoring, time-series network parameters are collected to complete feature extraction and model training, and link blocking risk status is predicted in real time. Analyze historical session thread execution data, integrate model prediction metrics, and construct a thread pool blocking profile for the local data region. Based on the quantitative statistical analysis of historical portrait samples, feature weights are adaptively calculated to complete the matching and retrieval of real-time portraits and historical portrait sets. The optimal segmentation optimistic waiting time is iteratively updated by combining the probability of incomplete generation. The system performs quantitative statistics on blocking across multiple regions and aggregates blocking features at the merchant level. It also calculates blocking feature parameters for the merchant's comprehensive incomplete bitstream and generates differentiated message distribution offset parameters to regulate asynchronous image processing and message routing partitions.

2. The method for cross-domain image anti-blocking asynchronous processing based on temporal supervised learning according to claim 1, characterized in that: The collected full-domain historical session logs are used to calculate the sharded optimistic wait times through hierarchical weighted fusion, including: Set up a synchronization session lifecycle monitoring component, call the distributed KV storage data observation interface to collect the full lifecycle data of cross-region synchronization sessions of images in each data area. The data area is a business storage area equipped with an independent storage cluster, a dedicated transmission line and a unique area identifier. Allocate an independent sharding waiting timer period for each synchronization session. The historical logs of each region are analyzed to separate the local data region and the synchronization session duration data of the other data regions. The session mean of the local data region and the session mean of all other data regions are calculated respectively. The average session duration difference is calculated in turn to obtain the regional latency deviation characteristics. The second local baseline waiting time is generated by combining the session mean of all other data regions. The ratio of the total number of session samples in the local data region to that in all other data regions is used as the fusion weight. The fusion weight is used to calculate the optimistic waiting time of the two types of local baseline waiting times and output the sharded optimistic waiting time. The waiting time is then corrected by taking into account the regional latency deviation to obtain the sharded optimistic waiting time.

3. The method for cross-domain image anti-blocking asynchronous processing based on temporal supervised learning according to claim 1, characterized in that: The real-time prediction of link congestion risk status also includes: Based on the session context instance container corresponding to each synchronization session, set network feature monitoring points, collect complete time-series operation data of the historical synchronization sessions in the local data area, and extract time-series network parameters such as instantaneous synchronization throughput, average transmission rate, leased line packet loss rate, and TCP retransmission count. Four types of temporal network parameters are selected as model input features. The fragmented state of historical sessions, KV lock spin time, task queue backlog, and fragmented bitstream generation are selected as sample annotation data to construct a temporal training sample set. Standardization preprocessing and temporal alignment processing are performed on the temporal training sample set. The model is iteratively converged through temporal supervised training and the model is solidified and stored. The model is an asynchronous image processing blocking prediction model. Within the time-based duration corresponding to the optimistic waiting time of the segment, four types of time-series network parameters of the cross-regional link are collected in real time based on the network feature monitoring points and input into the image asynchronous processing blocking prediction model. After time-series feature inference and calculation, the structured prediction data corresponding to the incomplete generation probability value, the expected total spin time of the KV batch read lock, the expected synchronous task queue accumulation length, and the expected incomplete code stream message increment corresponding to the future preset time period are output to obtain the real-time predicted link blocking risk status.

4. The method for cross-domain image anti-blocking asynchronous processing based on temporal supervised learning according to claim 1, characterized in that: The construction of the local data region thread pool blocking profile includes: Read the historical session execution logs stored in the local data area, extract the thread pool execution status data, and divide the thread objects into two categories: active worker threads and blocked worker threads. Statistically analyze the proportion of active worker threads, the increase in the number of blocked worker threads, the cumulative number of zombie threads, and the duration of thread starvation within a fixed session period to generate a multi-dimensional basic feature dataset for the thread pool. The model outputs four types of prediction indicators to retrieve all historical thread basic feature samples, filters historical samples that meet the matching requirements, and extracts four thread running parameters from the sample records. The arithmetic mean of the parameters of multiple matching samples is calculated to complete the aggregation. The four means are combined into a feature vector in a fixed order. The thread pool blocking profile of the local data area is generated based on the feature vector.

5. The method for cross-domain image anti-blocking asynchronous processing based on temporal supervised learning according to claim 1, characterized in that: The step of combining the incomplete generation probability to complete the optimal sharding optimistic waiting time iteration update includes: Read all historical thread pool blocking profiles persistently stored in the local data area to form a historical profile set. Then, perform a four-dimensional feature difference comparison between the real-time generated thread pool blocking profile and all profiles in the historical profile set. The four-dimensional features are the active worker thread occupancy ratio, the number of blocked worker threads, the cumulative number of zombie threads, and the duration of thread starvation. Traverse all samples in the historical portrait set, and for a single feature dimension, calculate the numerical dispersion and cumulative changes of thread blocking increment, incomplete bit stream generation increment, and thread resource leakage increment generated by synchronization. Accumulate the three types of increments to obtain the cumulative value of single-dimensional anomaly association. Determine the feature weight of each dimension based on the proportion of the cumulative value of all-dimensional anomaly association. Substitute the feature weights of each dimension to calculate the global matching degree between the real-time image and each historical image, filter the historical image with the highest global matching degree, and read the historical best shard waiting time parameter bound to the image.

6. The method for cross-domain image anti-blocking asynchronous processing based on temporal supervised learning according to claim 5, characterized in that: The method of combining the incomplete generation probability to complete the optimal slice optimistic waiting time iteration update also includes: The model's real-time output of the probability value of incomplete generation is retrieved. Using the historical best sharding waiting time as the base, the product of the base and the probability value of incomplete generation is added to obtain a new sharding waiting time value. The new sharding waiting time value is used to cover the original optimistic sharding waiting time. Two sets of mutual exclusion judgment logic are set up: when the probability value of incomplete generation is lower than the fixed statistical threshold and the cumulative number of zombie threads in the corresponding historical profile is zero, the historical best shard waiting time is retained; when the probability value of incomplete generation is greater than or equal to the fixed statistical threshold or the cumulative number of zombie threads in the corresponding historical profile is greater than zero, the optimistic waiting time of the current shard is updated by numerical superposition.

7. The method for cross-domain image anti-blocking asynchronous processing based on temporal supervised learning according to claim 1, characterized in that: The generation of differentiated message distribution offset parameters to regulate the asynchronous image processing message routing partitions includes: The optimistic waiting time of the updated shard is selected as the benchmark to continuously collect the thread pool operation characteristics and link operation characteristics within each waiting window. The frequency of incomplete shard generation, the cumulative number of spin-blocking threads, and the duration of continuous task queue accumulation within the window are statistically analyzed. The incomplete bitstream blocking characteristic parameters of the local data area are obtained through integrated calculation. At the same time, the thread pool operation data and incomplete bitstream generation data of all other data areas are collected. The three types of indicators of the corresponding areas are statistically analyzed and independently calculated to obtain the incomplete bitstream blocking characteristic parameters of the other data areas. A weighted aggregation operation is performed on the incomplete bitstream blocking feature parameters of the local data area and all other data areas to generate comprehensive incomplete bitstream blocking feature parameters of the entire domain; the merchant's unique identifier is selected as the grouping condition to isolate and collect cross-region asynchronous processing tasks of product images across the entire domain, and the total number of cross-region synchronous tasks, the total number of incomplete bitstreams generated, the frequency of thread blocking triggers, and the average backlog time of the task queue are statistically analyzed for a single merchant. Four statistical indicators for a single merchant are selected and subjected to global extreme value normalization in sequence to map all indicators to a unified value range; fixed statistical weights are assigned to the historical abnormal contribution ratios of the four indicators, and the four normalized and weighted indicators are summed to output the merchant's comprehensive incomplete bitstream blocking feature parameters. The numerical range of the comprehensive incomplete bitstream blocking feature parameters of each merchant is selected to generate the corresponding message distribution offset parameters. The hash route offset operation is performed on the asynchronous processing message of the product image generated by the fragment merging. The routing offset is dynamically adjusted according to the merchant blocking feature. Merchant messages corresponding to high blocking features are assigned to multiple consumption partitions in the whole domain, and merchant messages corresponding to low blocking features are assigned to idle consumption partitions in the whole domain.

8. A cross-domain image anti-blocking asynchronous processing system based on temporal supervised learning, which is applied to the cross-domain image anti-blocking asynchronous processing method based on temporal supervised learning as described in any one of claims 1-7, characterized in that: The cross-domain image anti-blocking asynchronous processing system includes a timing control module, a risk prediction module, a profile construction module, and a routing scheduling module. The timing control module is used to collect synchronous session data, calculate and iteratively update the optimistic waiting time of the shards; the risk prediction module is used to collect link timing network parameters, train the blocking prediction model and output structured prediction data; the profile construction module is used to extract thread pool running features, construct thread pool blocking profiles and calculate feature weights; the routing scheduling module is used to statistically analyze the blocking feature parameters of the entire domain and merchant incomplete bitstreams and control the routing partitions of asynchronous image processing messages.

9. A cross-domain image anti-blocking asynchronous processing system based on temporal supervised learning according to claim 8, characterized in that: The timing control module includes a session monitoring unit and a duration calculation unit. The session monitoring unit is used to deploy a synchronous session lifecycle monitoring component, collect cross-regional synchronous session full-cycle data of product images in each data region, and establish a sharded optimistic waiting window timing mechanism. The duration calculation unit is used to parse the global historical session logs, calculate two types of local baseline waiting durations in a hierarchical manner, weightedly fuse and output the sharded optimistic waiting duration, and link the thread pool blocking profile to iteratively update the sharded optimistic waiting duration. The risk prediction module includes a link acquisition unit and a model inference unit. The link acquisition unit is used to deploy network feature monitoring points in the session context instance container and continuously collect link time-series network parameters. The model inference unit is used to construct a time-series training sample set to complete model training and output multiple types of structured blocking prediction data based on real-time link parameters.

10. A cross-domain image anti-blocking asynchronous processing system based on temporal supervised learning according to claim 8, characterized in that: The profile construction module includes a thread parsing unit and a feature matching unit. The thread parsing unit is used to read historical session running logs, extract thread pool running metrics, and generate a multi-dimensional basic feature dataset of the thread pool. The feature matching unit is used to retrieve and match historical thread samples to aggregate and generate feature vectors, construct a thread pool blocking profile in the local data area, quantitatively calculate the weight of each feature dimension, and complete the profile set matching retrieval. The routing scheduling module includes a global statistics unit and a message distribution unit. The global statistics unit is used to collect incomplete bitstream blocking feature parameters for each data region and aggregate merchant-level indicators to generate comprehensive incomplete bitstream blocking feature parameters for merchants. The message distribution unit is used to select comprehensive incomplete bitstream blocking feature parameters for merchants to generate message distribution offset parameters and perform hash routing offset calculations to adjust the asynchronous image processing message delivery and consumption partitions.

Images